***UNCHECKED*** SolrCloud is sick.

classic Classic list List threaded Threaded
40 messages Options
12
Reply | Threaded
Open this post in threaded view
|

***UNCHECKED*** SolrCloud is sick.

Mark Miller-3
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Scott Blum
Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).

What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?

On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Noble Paul നോബിള്‍  नोब्ळ्
Hi, 

I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:

* How does it work?

* What are the performance characteristics? Can it be done more efficiently?

* What are the public touch points?

** Which are the files we store in ZK? Are they expected to be watched always?

** Or are they read on demand?

** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?


We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt. 


This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.) 


Now, why am I discussing this in this thread?


While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.


I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion. 


I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
Thanks

On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).

What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?

On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
I am not personally attacking anyone.

Everyone here does good work in one way or another.

However, if you cross a bar of prolific vs attention to detail, tests, doc, and things beyond just you, I do hold you to a higher standard.

You are super prolific and your care for others following your footsteps or full consideration to things, is lacking. I see nothing personal about it, it’s a pure code and work observation.

You have good ideas, you can spin code, that’s awesome, but your often using that power in a way that clears a lot of ground without a lot of replacement tree planting.

I’ve said this 50 times, I’ve said this to the PMC more than once, it’s nothing personal. Personally you are a great human and enjoyable to interact with. I always enjoy you, my wife still tells me she enjoys you. 

I’ve got problems with others work as well, but no one else seems to have the same disregard for the other devs while working.

I wish Yonik would document more for mere mortals. I wish AB would look at more the costs he adds to startup and shutdown. I wish, I wish... I dint care as long as you have a cooperative attitude. 

And I can help with the area people are lacking.  It’s no free lunch, it’s a going to be a pain I. The ass, we can talk another that when I have my proposal planned. 

There is tons of stuff you can help with project with, prolific people are valuable. But we need to figure out a better way to operate and we need to start from more stable ground.

I’m not tossing some design on you. We have a design. It’s fine. Our impl and foundation are bad, they have always been bad, we are going backward faster than forward.

Most of the worst and basic code is mine.

We all write bad code. If you are open to owning that your efforts could be improved in various area - maybe even just to please the others we work with, that’s enough for me. It’s fine then. But an obstinate disregard for the fact that we all share this project ... I can’t work wile with it. 

Anyway, I’ve got stuff for us to build with. I’d rather you spend your time making things fast that we care about and not ghosts of the past.

I know everyone has wanted that. I’m not here saying you guys fucked up. I’m at the head of that list. It’s our project.

Mark

On Sat, Nov 2, 2019 at 6:32 PM Noble Paul <[hidden email]> wrote:
Hi, 

I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:

* How does it work?

* What are the performance characteristics? Can it be done more efficiently?

* What are the public touch points?

** Which are the files we store in ZK? Are they expected to be watched always?

** Or are they read on demand?

** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?


We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt. 


This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.) 


Now, why am I discussing this in this thread?


While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.


I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion. 


I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
Thanks

On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).

What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?

On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--
--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
In reply to this post by Scott Blum
Not much. Something you can understand. How about tests < 10 seconds fail or not. Good logging and as a backup good debug logging. Docs on how things are designed to work? Tracking of all important operations and how long they take with tight cutoffs? Proper response to interruption 100% of the time? The idea of a cluster start and stop? Of a cluster install to ZK initially. Drop all legacyCloud support, stateformat=1 support, maybe a few other things.

I've got some stuff, I'm gonna pull out as fast as I sensibly can given many setbacks and too little sleep for a long time.

I'm not here to do all the of the lift for everyone, but unless I get sick in the next week or two or my 10 backup methods and git pushes and backup branches fail or I just burn the hell out, I have a solid refuge that we can knock out and then build on with confidence.

- Mark

On Sat, Nov 2, 2019 at 5:52 PM Scott Blum <[hidden email]> wrote:
Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).

What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?

On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
I mean the reality is - why do we not have just a single watcher per node pulling in state. We are we not tracking and minimizing state transfers and changes? Why are we not measuring the time it takes to round trip a state.json and adjusting? Looking at load to adjust overseerish duties and leader election? A million other smart things?

Because it's too hard. It's too hard and we all gave up long ago on figuring out what to do about it. Because we are programming in assembly in an abyss when we should be doing java in the clouds.

Everyone knows the SolrCloud DNA one way or another.We all somehow made our peace with it or not.

It's easy when you dont go deep. Hell thats easy to forget even if you do.

But I'm looping on it now, have to eject.

- Mark

On Sat, Nov 2, 2019 at 10:15 PM Mark Miller <[hidden email]> wrote:
Not much. Something you can understand. How about tests < 10 seconds fail or not. Good logging and as a backup good debug logging. Docs on how things are designed to work? Tracking of all important operations and how long they take with tight cutoffs? Proper response to interruption 100% of the time? The idea of a cluster start and stop? Of a cluster install to ZK initially. Drop all legacyCloud support, stateformat=1 support, maybe a few other things.

I've got some stuff, I'm gonna pull out as fast as I sensibly can given many setbacks and too little sleep for a long time.

I'm not here to do all the of the lift for everyone, but unless I get sick in the next week or two or my 10 backup methods and git pushes and backup branches fail or I just burn the hell out, I have a solid refuge that we can knock out and then build on with confidence.

- Mark

On Sat, Nov 2, 2019 at 5:52 PM Scott Blum <[hidden email]> wrote:
Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).

What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?

On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--


--


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
And it didnt get any easier. What I did about it is kill myself multiple times over 2 years for weeks on end of torturing my wife. And I found a million problems, a million bugs, a million terrible inefficiencies. And I fixed and lost countless of them friggen twice. And didnt lose tons of the work as well. And so it's not easy to get out of this. Its not easy at all. And i havent even done the hard part yet.

- Mark

On Sat, Nov 2, 2019 at 10:24 PM Mark Miller <[hidden email]> wrote:
I mean the reality is - why do we not have just a single watcher per node pulling in state. We are we not tracking and minimizing state transfers and changes? Why are we not measuring the time it takes to round trip a state.json and adjusting? Looking at load to adjust overseerish duties and leader election? A million other smart things?

Because it's too hard. It's too hard and we all gave up long ago on figuring out what to do about it. Because we are programming in assembly in an abyss when we should be doing java in the clouds.

Everyone knows the SolrCloud DNA one way or another.We all somehow made our peace with it or not.

It's easy when you dont go deep. Hell thats easy to forget even if you do.

But I'm looping on it now, have to eject.

- Mark

On Sat, Nov 2, 2019 at 10:15 PM Mark Miller <[hidden email]> wrote:
Not much. Something you can understand. How about tests < 10 seconds fail or not. Good logging and as a backup good debug logging. Docs on how things are designed to work? Tracking of all important operations and how long they take with tight cutoffs? Proper response to interruption 100% of the time? The idea of a cluster start and stop? Of a cluster install to ZK initially. Drop all legacyCloud support, stateformat=1 support, maybe a few other things.

I've got some stuff, I'm gonna pull out as fast as I sensibly can given many setbacks and too little sleep for a long time.

I'm not here to do all the of the lift for everyone, but unless I get sick in the next week or two or my 10 backup methods and git pushes and backup branches fail or I just burn the hell out, I have a solid refuge that we can knock out and then build on with confidence.

- Mark

On Sat, Nov 2, 2019 at 5:52 PM Scott Blum <[hidden email]> wrote:
Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).

What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?

On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--


--


--


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
Things are also counterintuitive. The more you fix and the faster things work the more things fail. It’s like rings of hell. 

Mark

On Sat, Nov 2, 2019 at 10:29 PM Mark Miller <[hidden email]> wrote:
And it didnt get any easier. What I did about it is kill myself multiple times over 2 years for weeks on end of torturing my wife. And I found a million problems, a million bugs, a million terrible inefficiencies. And I fixed and lost countless of them friggen twice. And didnt lose tons of the work as well. And so it's not easy to get out of this. Its not easy at all. And i havent even done the hard part yet.

- Mark

On Sat, Nov 2, 2019 at 10:24 PM Mark Miller <[hidden email]> wrote:
I mean the reality is - why do we not have just a single watcher per node pulling in state. We are we not tracking and minimizing state transfers and changes? Why are we not measuring the time it takes to round trip a state.json and adjusting? Looking at load to adjust overseerish duties and leader election? A million other smart things?

Because it's too hard. It's too hard and we all gave up long ago on figuring out what to do about it. Because we are programming in assembly in an abyss when we should be doing java in the clouds.

Everyone knows the SolrCloud DNA one way or another.We all somehow made our peace with it or not.

It's easy when you dont go deep. Hell thats easy to forget even if you do.

But I'm looping on it now, have to eject.

- Mark

On Sat, Nov 2, 2019 at 10:15 PM Mark Miller <[hidden email]> wrote:
Not much. Something you can understand. How about tests < 10 seconds fail or not. Good logging and as a backup good debug logging. Docs on how things are designed to work? Tracking of all important operations and how long they take with tight cutoffs? Proper response to interruption 100% of the time? The idea of a cluster start and stop? Of a cluster install to ZK initially. Drop all legacyCloud support, stateformat=1 support, maybe a few other things.

I've got some stuff, I'm gonna pull out as fast as I sensibly can given many setbacks and too little sleep for a long time.

I'm not here to do all the of the lift for everyone, but unless I get sick in the next week or two or my 10 backup methods and git pushes and backup branches fail or I just burn the hell out, I have a solid refuge that we can knock out and then build on with confidence.

- Mark

On Sat, Nov 2, 2019 at 5:52 PM Scott Blum <[hidden email]> wrote:
Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).

What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?

On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--


--


--


--
--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

david.w.smiley@gmail.com
In reply to this post by Noble Paul നോബിള്‍ नोब्ळ्
Yeah we do a bad job of the things you listed Noble.  :-(   My colleagues want pointers to internal docs but the sad reality is there isn't any.  You may notice I'm a stickler in my code reviews for requiring javadocs on all top level classes.  I think more javadocs and code comments would be very helpful -- especially for the major classes.  This might help us all and others a lot more.  For example I think Lucene does a rather fine job of this for its major classes -- IndexWriter being a good example.

~ David Smiley
Apache Lucene/Solr Search Developer


On Sat, Nov 2, 2019 at 7:32 PM Noble Paul <[hidden email]> wrote:
Hi, 

I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:

* How does it work?

* What are the performance characteristics? Can it be done more efficiently?

* What are the public touch points?

** Which are the files we store in ZK? Are they expected to be watched always?

** Or are they read on demand?

** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?


We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt. 


This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.) 


Now, why am I discussing this in this thread?


While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.


I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion. 


I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
Thanks

On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).

What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?

On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Noble Paul നോബിള്‍  नोब्ळ्
Solr has to do more than Lucene. A Lucene user is mostly a developer
who reads javadocs. A Solr user's touch points are

* Public API
* Ref guide
* publicly visible files (in ZK as well as file system)
* What to see/look for in the log files to debug issues

Then we have more nuanced touch points such as the knowledge base of
what happens internally in the system when 'X' API is invoked or when
'Y' behavior is observed in ZK data.

The problem with delaying the review process till code completion is
that, any changes based on review comments will require massive amount
of work.

I don't have an answer to how we achieve it. But, I clearly see this
as a major gap in our development process today.

This discussion may not be relevant in this thread, may be because no
behavior is changed at all. We don't know yet

What I want to believe is Mark is doing the right thing & it's gonna
help us all in dealing with our operational issues. I don't want to
interrupt his work with more discussions.

Thanks you


On Sun, Nov 3, 2019 at 3:32 PM David Smiley <[hidden email]> wrote:

>
> Yeah we do a bad job of the things you listed Noble.  :-(   My colleagues want pointers to internal docs but the sad reality is there isn't any.  You may notice I'm a stickler in my code reviews for requiring javadocs on all top level classes.  I think more javadocs and code comments would be very helpful -- especially for the major classes.  This might help us all and others a lot more.  For example I think Lucene does a rather fine job of this for its major classes -- IndexWriter being a good example.
>
> ~ David Smiley
> Apache Lucene/Solr Search Developer
> http://www.linkedin.com/in/davidwsmiley
>
>
> On Sat, Nov 2, 2019 at 7:32 PM Noble Paul <[hidden email]> wrote:
>>
>> Hi,
>>
>> I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:
>>
>> * How does it work?
>>
>> * What are the performance characteristics? Can it be done more efficiently?
>>
>> * What are the public touch points?
>>
>> ** Which are the files we store in ZK? Are they expected to be watched always?
>>
>> ** Or are they read on demand?
>>
>> ** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?
>>
>>
>> We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt.
>>
>>
>> This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.)
>>
>>
>> Now, why am I discussing this in this thread?
>>
>>
>> While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.
>>
>>
>> I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion.
>>
>>
>> I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
>> Thanks
>>
>> On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
>>>
>>> Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).
>>>
>>> What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?
>>>
>>> On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
>>>>
>>>> Give me a short bit to follow up and I will lay out my case and proposal.
>>>>
>>>> Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.
>>>>
>>>> If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.
>>>>
>>>> - mark
>>>>
>>>> On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
>>>>>
>>>>> Hey All Solr Dev's,
>>>>>
>>>>> SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).
>>>>>
>>>>> Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.
>>>>>
>>>>> This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.
>>>>>
>>>>> There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.
>>>>>
>>>>> So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.
>>>>>
>>>>> We should all be embarrassed by the state of things.
>>>>>
>>>>> I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.
>>>>>
>>>>> --
>>>>> - Mark
>>>>>
>>>>> http://about.me/markrmiller
>>>>
>>>>
>>>>
>>>> --
>>>> - Mark
>>>>
>>>> http://about.me/markrmiller



--
-----------------------------------------------------
Noble Paul

---------------------------------------------------------------------
To unsubscribe, e-mail: [hidden email]
For additional commands, e-mail: [hidden email]

Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Jörn Franke
I cannot say anything about the statements, but maybe it could help to introduce Solr Improvement Proposals (SIP) similar to Kafka Improvement Proposals (KIP) or Flink Improvement Proposals (FLIP).

  I think they are helpful to facilitate design decisions and refactoring / redesign decision. They also would allow it to do it in an iterative manner without changing everything at once.
The final version could be out  in The Git of Solr in markdown including figures presenting parts of the design.

However for developing them I propose a more inclusive approach where many people (not only core developers) can easily comment and support, eg Google docs or similar.

> Am 03.11.2019 um 06:39 schrieb Noble Paul <[hidden email]>:
>
> Solr has to do more than Lucene. A Lucene user is mostly a developer
> who reads javadocs. A Solr user's touch points are
>
> * Public API
> * Ref guide
> * publicly visible files (in ZK as well as file system)
> * What to see/look for in the log files to debug issues
>
> Then we have more nuanced touch points such as the knowledge base of
> what happens internally in the system when 'X' API is invoked or when
> 'Y' behavior is observed in ZK data.
>
> The problem with delaying the review process till code completion is
> that, any changes based on review comments will require massive amount
> of work.
>
> I don't have an answer to how we achieve it. But, I clearly see this
> as a major gap in our development process today.
>
> This discussion may not be relevant in this thread, may be because no
> behavior is changed at all. We don't know yet
>
> What I want to believe is Mark is doing the right thing & it's gonna
> help us all in dealing with our operational issues. I don't want to
> interrupt his work with more discussions.
>
> Thanks you
>
>
>> On Sun, Nov 3, 2019 at 3:32 PM David Smiley <[hidden email]> wrote:
>>
>> Yeah we do a bad job of the things you listed Noble.  :-(   My colleagues want pointers to internal docs but the sad reality is there isn't any.  You may notice I'm a stickler in my code reviews for requiring javadocs on all top level classes.  I think more javadocs and code comments would be very helpful -- especially for the major classes.  This might help us all and others a lot more.  For example I think Lucene does a rather fine job of this for its major classes -- IndexWriter being a good example.
>>
>> ~ David Smiley
>> Apache Lucene/Solr Search Developer
>> http://www.linkedin.com/in/davidwsmiley
>>
>>
>>> On Sat, Nov 2, 2019 at 7:32 PM Noble Paul <[hidden email]> wrote:
>>>
>>> Hi,
>>>
>>> I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:
>>>
>>> * How does it work?
>>>
>>> * What are the performance characteristics? Can it be done more efficiently?
>>>
>>> * What are the public touch points?
>>>
>>> ** Which are the files we store in ZK? Are they expected to be watched always?
>>>
>>> ** Or are they read on demand?
>>>
>>> ** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?
>>>
>>>
>>> We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt.
>>>
>>>
>>> This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.)
>>>
>>>
>>> Now, why am I discussing this in this thread?
>>>
>>>
>>> While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.
>>>
>>>
>>> I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion.
>>>
>>>
>>> I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
>>> Thanks
>>>
>>> On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
>>>>
>>>> Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).
>>>>
>>>> What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?
>>>>
>>>> On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
>>>>>
>>>>> Give me a short bit to follow up and I will lay out my case and proposal.
>>>>>
>>>>> Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.
>>>>>
>>>>> If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.
>>>>>
>>>>> - mark
>>>>>
>>>>> On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
>>>>>>
>>>>>> Hey All Solr Dev's,
>>>>>>
>>>>>> SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).
>>>>>>
>>>>>> Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.
>>>>>>
>>>>>> This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.
>>>>>>
>>>>>> There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.
>>>>>>
>>>>>> So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.
>>>>>>
>>>>>> We should all be embarrassed by the state of things.
>>>>>>
>>>>>> I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.
>>>>>>
>>>>>> --
>>>>>> - Mark
>>>>>>
>>>>>> http://about.me/markrmiller
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> - Mark
>>>>>
>>>>> http://about.me/markrmiller
>
>
>
> --
> -----------------------------------------------------
> Noble Paul
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: [hidden email]
> For additional commands, e-mail: [hidden email]
>

---------------------------------------------------------------------
To unsubscribe, e-mail: [hidden email]
For additional commands, e-mail: [hidden email]

Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
bq.  They also would allow it to do it in an iterative manner without changing everything at once.

Sadly, you can't fix this piece by piece :) I dare anyone to try. I encourage, I applaud the effort.

The world is your oyster from a good spot - take your pick of how to do things.

But from this spot, if anyone thinks we are getting out design change by design change, JIRA by JIRA, I'm so sorry. Let's commiserate in a couple years on a beer when you  give up on that.

- Mark

On Sun, Nov 3, 2019 at 4:01 AM Jörn Franke <[hidden email]> wrote:
I cannot say anything about the statements, but maybe it could help to introduce Solr Improvement Proposals (SIP) similar to Kafka Improvement Proposals (KIP) or Flink Improvement Proposals (FLIP).

  I think they are helpful to facilitate design decisions and refactoring / redesign decision. They also would allow it to do it in an iterative manner without changing everything at once.
The final version could be out  in The Git of Solr in markdown including figures presenting parts of the design.

However for developing them I propose a more inclusive approach where many people (not only core developers) can easily comment and support, eg Google docs or similar.

> Am 03.11.2019 um 06:39 schrieb Noble Paul <[hidden email]>:
>
> Solr has to do more than Lucene. A Lucene user is mostly a developer
> who reads javadocs. A Solr user's touch points are
>
> * Public API
> * Ref guide
> * publicly visible files (in ZK as well as file system)
> * What to see/look for in the log files to debug issues
>
> Then we have more nuanced touch points such as the knowledge base of
> what happens internally in the system when 'X' API is invoked or when
> 'Y' behavior is observed in ZK data.
>
> The problem with delaying the review process till code completion is
> that, any changes based on review comments will require massive amount
> of work.
>
> I don't have an answer to how we achieve it. But, I clearly see this
> as a major gap in our development process today.
>
> This discussion may not be relevant in this thread, may be because no
> behavior is changed at all. We don't know yet
>
> What I want to believe is Mark is doing the right thing & it's gonna
> help us all in dealing with our operational issues. I don't want to
> interrupt his work with more discussions.
>
> Thanks you
>
>
>> On Sun, Nov 3, 2019 at 3:32 PM David Smiley <[hidden email]> wrote:
>>
>> Yeah we do a bad job of the things you listed Noble.  :-(   My colleagues want pointers to internal docs but the sad reality is there isn't any.  You may notice I'm a stickler in my code reviews for requiring javadocs on all top level classes.  I think more javadocs and code comments would be very helpful -- especially for the major classes.  This might help us all and others a lot more.  For example I think Lucene does a rather fine job of this for its major classes -- IndexWriter being a good example.
>>
>> ~ David Smiley
>> Apache Lucene/Solr Search Developer
>> http://www.linkedin.com/in/davidwsmiley
>>
>>
>>> On Sat, Nov 2, 2019 at 7:32 PM Noble Paul <[hidden email]> wrote:
>>>
>>> Hi,
>>>
>>> I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:
>>>
>>> * How does it work?
>>>
>>> * What are the performance characteristics? Can it be done more efficiently?
>>>
>>> * What are the public touch points?
>>>
>>> ** Which are the files we store in ZK? Are they expected to be watched always?
>>>
>>> ** Or are they read on demand?
>>>
>>> ** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?
>>>
>>>
>>> We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt.
>>>
>>>
>>> This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.)
>>>
>>>
>>> Now, why am I discussing this in this thread?
>>>
>>>
>>> While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.
>>>
>>>
>>> I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion.
>>>
>>>
>>> I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
>>> Thanks
>>>
>>> On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
>>>>
>>>> Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).
>>>>
>>>> What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?
>>>>
>>>> On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
>>>>>
>>>>> Give me a short bit to follow up and I will lay out my case and proposal.
>>>>>
>>>>> Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.
>>>>>
>>>>> If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.
>>>>>
>>>>> - mark
>>>>>
>>>>> On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
>>>>>>
>>>>>> Hey All Solr Dev's,
>>>>>>
>>>>>> SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).
>>>>>>
>>>>>> Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.
>>>>>>
>>>>>> This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.
>>>>>>
>>>>>> There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.
>>>>>>
>>>>>> So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.
>>>>>>
>>>>>> We should all be embarrassed by the state of things.
>>>>>>
>>>>>> I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.
>>>>>>
>>>>>> --
>>>>>> - Mark
>>>>>>
>>>>>> http://about.me/markrmiller
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> - Mark
>>>>>
>>>>> http://about.me/markrmiller
>
>
>
> --
> -----------------------------------------------------
> Noble Paul
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: [hidden email]
> For additional commands, e-mail: [hidden email]
>

---------------------------------------------------------------------
To unsubscribe, e-mail: [hidden email]
For additional commands, e-mail: [hidden email]



--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
In fact this will be a fundamental difference some of us are about to split between.

Those that think they can ever fix the tests or the system or the 1000s of bugs we have and keep adding due to our current world view of making tests fit the system not the system fit the tests and that fact that everything is so slow and retry and workaround that stupid shit works all over. It's all deep. It's ingrained. It grown over for a decade.Its a project of 60 modules.

Soon we will split between those that think they are making progress across the ocean and those that think we are sitting in shark infested waters waiting to die actually, starting to float backwards sometimes now.

- Mark

On Sun, Nov 3, 2019 at 7:23 AM Mark Miller <[hidden email]> wrote:
bq.  They also would allow it to do it in an iterative manner without changing everything at once.

Sadly, you can't fix this piece by piece :) I dare anyone to try. I encourage, I applaud the effort.

The world is your oyster from a good spot - take your pick of how to do things.

But from this spot, if anyone thinks we are getting out design change by design change, JIRA by JIRA, I'm so sorry. Let's commiserate in a couple years on a beer when you  give up on that.

- Mark

On Sun, Nov 3, 2019 at 4:01 AM Jörn Franke <[hidden email]> wrote:
I cannot say anything about the statements, but maybe it could help to introduce Solr Improvement Proposals (SIP) similar to Kafka Improvement Proposals (KIP) or Flink Improvement Proposals (FLIP).

  I think they are helpful to facilitate design decisions and refactoring / redesign decision. They also would allow it to do it in an iterative manner without changing everything at once.
The final version could be out  in The Git of Solr in markdown including figures presenting parts of the design.

However for developing them I propose a more inclusive approach where many people (not only core developers) can easily comment and support, eg Google docs or similar.

> Am 03.11.2019 um 06:39 schrieb Noble Paul <[hidden email]>:
>
> Solr has to do more than Lucene. A Lucene user is mostly a developer
> who reads javadocs. A Solr user's touch points are
>
> * Public API
> * Ref guide
> * publicly visible files (in ZK as well as file system)
> * What to see/look for in the log files to debug issues
>
> Then we have more nuanced touch points such as the knowledge base of
> what happens internally in the system when 'X' API is invoked or when
> 'Y' behavior is observed in ZK data.
>
> The problem with delaying the review process till code completion is
> that, any changes based on review comments will require massive amount
> of work.
>
> I don't have an answer to how we achieve it. But, I clearly see this
> as a major gap in our development process today.
>
> This discussion may not be relevant in this thread, may be because no
> behavior is changed at all. We don't know yet
>
> What I want to believe is Mark is doing the right thing & it's gonna
> help us all in dealing with our operational issues. I don't want to
> interrupt his work with more discussions.
>
> Thanks you
>
>
>> On Sun, Nov 3, 2019 at 3:32 PM David Smiley <[hidden email]> wrote:
>>
>> Yeah we do a bad job of the things you listed Noble.  :-(   My colleagues want pointers to internal docs but the sad reality is there isn't any.  You may notice I'm a stickler in my code reviews for requiring javadocs on all top level classes.  I think more javadocs and code comments would be very helpful -- especially for the major classes.  This might help us all and others a lot more.  For example I think Lucene does a rather fine job of this for its major classes -- IndexWriter being a good example.
>>
>> ~ David Smiley
>> Apache Lucene/Solr Search Developer
>> http://www.linkedin.com/in/davidwsmiley
>>
>>
>>> On Sat, Nov 2, 2019 at 7:32 PM Noble Paul <[hidden email]> wrote:
>>>
>>> Hi,
>>>
>>> I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:
>>>
>>> * How does it work?
>>>
>>> * What are the performance characteristics? Can it be done more efficiently?
>>>
>>> * What are the public touch points?
>>>
>>> ** Which are the files we store in ZK? Are they expected to be watched always?
>>>
>>> ** Or are they read on demand?
>>>
>>> ** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?
>>>
>>>
>>> We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt.
>>>
>>>
>>> This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.)
>>>
>>>
>>> Now, why am I discussing this in this thread?
>>>
>>>
>>> While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.
>>>
>>>
>>> I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion.
>>>
>>>
>>> I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
>>> Thanks
>>>
>>> On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
>>>>
>>>> Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).
>>>>
>>>> What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?
>>>>
>>>> On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
>>>>>
>>>>> Give me a short bit to follow up and I will lay out my case and proposal.
>>>>>
>>>>> Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.
>>>>>
>>>>> If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.
>>>>>
>>>>> - mark
>>>>>
>>>>> On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
>>>>>>
>>>>>> Hey All Solr Dev's,
>>>>>>
>>>>>> SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).
>>>>>>
>>>>>> Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.
>>>>>>
>>>>>> This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.
>>>>>>
>>>>>> There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.
>>>>>>
>>>>>> So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.
>>>>>>
>>>>>> We should all be embarrassed by the state of things.
>>>>>>
>>>>>> I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.
>>>>>>
>>>>>> --
>>>>>> - Mark
>>>>>>
>>>>>> http://about.me/markrmiller
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> - Mark
>>>>>
>>>>> http://about.me/markrmiller
>
>
>
> --
> -----------------------------------------------------
> Noble Paul
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: [hidden email]
> For additional commands, e-mail: [hidden email]
>

---------------------------------------------------------------------
To unsubscribe, e-mail: [hidden email]
For additional commands, e-mail: [hidden email]



--


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
Personally, I believe the latter so strongly, if I can’t convince the others in the raft with me, I’m jumping in and swimming to another raft after my entire adult life here. 

Mark

On Sun, Nov 3, 2019 at 7:30 AM Mark Miller <[hidden email]> wrote:
In fact this will be a fundamental difference some of us are about to split between.

Those that think they can ever fix the tests or the system or the 1000s of bugs we have and keep adding due to our current world view of making tests fit the system not the system fit the tests and that fact that everything is so slow and retry and workaround that stupid shit works all over. It's all deep. It's ingrained. It grown over for a decade.Its a project of 60 modules.

Soon we will split between those that think they are making progress across the ocean and those that think we are sitting in shark infested waters waiting to die actually, starting to float backwards sometimes now.

- Mark

On Sun, Nov 3, 2019 at 7:23 AM Mark Miller <[hidden email]> wrote:
bq.  They also would allow it to do it in an iterative manner without changing everything at once.

Sadly, you can't fix this piece by piece :) I dare anyone to try. I encourage, I applaud the effort.

The world is your oyster from a good spot - take your pick of how to do things.

But from this spot, if anyone thinks we are getting out design change by design change, JIRA by JIRA, I'm so sorry. Let's commiserate in a couple years on a beer when you  give up on that.

- Mark

On Sun, Nov 3, 2019 at 4:01 AM Jörn Franke <[hidden email]> wrote:
I cannot say anything about the statements, but maybe it could help to introduce Solr Improvement Proposals (SIP) similar to Kafka Improvement Proposals (KIP) or Flink Improvement Proposals (FLIP).

  I think they are helpful to facilitate design decisions and refactoring / redesign decision. They also would allow it to do it in an iterative manner without changing everything at once.
The final version could be out  in The Git of Solr in markdown including figures presenting parts of the design.

However for developing them I propose a more inclusive approach where many people (not only core developers) can easily comment and support, eg Google docs or similar.

> Am 03.11.2019 um 06:39 schrieb Noble Paul <[hidden email]>:
>
> Solr has to do more than Lucene. A Lucene user is mostly a developer
> who reads javadocs. A Solr user's touch points are
>
> * Public API
> * Ref guide
> * publicly visible files (in ZK as well as file system)
> * What to see/look for in the log files to debug issues
>
> Then we have more nuanced touch points such as the knowledge base of
> what happens internally in the system when 'X' API is invoked or when
> 'Y' behavior is observed in ZK data.
>
> The problem with delaying the review process till code completion is
> that, any changes based on review comments will require massive amount
> of work.
>
> I don't have an answer to how we achieve it. But, I clearly see this
> as a major gap in our development process today.
>
> This discussion may not be relevant in this thread, may be because no
> behavior is changed at all. We don't know yet
>
> What I want to believe is Mark is doing the right thing & it's gonna
> help us all in dealing with our operational issues. I don't want to
> interrupt his work with more discussions.
>
> Thanks you
>
>
>> On Sun, Nov 3, 2019 at 3:32 PM David Smiley <[hidden email]> wrote:
>>
>> Yeah we do a bad job of the things you listed Noble.  :-(   My colleagues want pointers to internal docs but the sad reality is there isn't any.  You may notice I'm a stickler in my code reviews for requiring javadocs on all top level classes.  I think more javadocs and code comments would be very helpful -- especially for the major classes.  This might help us all and others a lot more.  For example I think Lucene does a rather fine job of this for its major classes -- IndexWriter being a good example.
>>
>> ~ David Smiley
>> Apache Lucene/Solr Search Developer
>> http://www.linkedin.com/in/davidwsmiley
>>
>>
>>> On Sat, Nov 2, 2019 at 7:32 PM Noble Paul <[hidden email]> wrote:
>>>
>>> Hi,
>>>
>>> I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:
>>>
>>> * How does it work?
>>>
>>> * What are the performance characteristics? Can it be done more efficiently?
>>>
>>> * What are the public touch points?
>>>
>>> ** Which are the files we store in ZK? Are they expected to be watched always?
>>>
>>> ** Or are they read on demand?
>>>
>>> ** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?
>>>
>>>
>>> We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt.
>>>
>>>
>>> This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.)
>>>
>>>
>>> Now, why am I discussing this in this thread?
>>>
>>>
>>> While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.
>>>
>>>
>>> I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion.
>>>
>>>
>>> I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
>>> Thanks
>>>
>>> On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
>>>>
>>>> Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).
>>>>
>>>> What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?
>>>>
>>>> On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
>>>>>
>>>>> Give me a short bit to follow up and I will lay out my case and proposal.
>>>>>
>>>>> Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.
>>>>>
>>>>> If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.
>>>>>
>>>>> - mark
>>>>>
>>>>> On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
>>>>>>
>>>>>> Hey All Solr Dev's,
>>>>>>
>>>>>> SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).
>>>>>>
>>>>>> Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.
>>>>>>
>>>>>> This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.
>>>>>>
>>>>>> There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.
>>>>>>
>>>>>> So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.
>>>>>>
>>>>>> We should all be embarrassed by the state of things.
>>>>>>
>>>>>> I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.
>>>>>>
>>>>>> --
>>>>>> - Mark
>>>>>>
>>>>>> http://about.me/markrmiller
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> - Mark
>>>>>
>>>>> http://about.me/markrmiller
>
>
>
> --
> -----------------------------------------------------
> Noble Paul
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: [hidden email]
> For additional commands, e-mail: [hidden email]
>

---------------------------------------------------------------------
To unsubscribe, e-mail: [hidden email]
For additional commands, e-mail: [hidden email]



--


--
--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
From a credentials standpoint:

Yonik and I built 90% of it originally and then I spent years on it with few other devs or users.

Pretty sure I'm the only one that has ever had 95%+ of the Solr test suites work in under 10-15 seconds consistently - 4000 tests across like 1000 suites. Got them all to run in parallel in under 5 minutes vs the 20-45 it takes on a good day after tons of other speed ups and fixes I've already done.

I doubt there are many insane enough that have deep dived and pushed around the entire code base for 2-3 weeks multiple times, 16-20 hours a day. That has spent the last decade beyond that stupid time almost exclusively on this system. Designing it with Yonik, building it, fixing it, helping people with it, monitoring it, responding to pages and escalations for it.

That has spent half a year replacing the entire decade old build for Lucene and Solr and all it's various nooks and crannies.

If anyone has spent more time on this system or pushing it around on a large scale or or has seen it in 100x the shape it is now more than once, please speak up, you are in charge, I follow you. I'm not that bright, if you have done the ground work, default to you.

Otherwise, I don't even have much confidence anyone else even knows this system remotely well. All that time and effort and the most I know of it is what awful awful shape its in and the bad trend direction.


- Mark

On Sun, Nov 3, 2019 at 7:35 AM Mark Miller <[hidden email]> wrote:
Personally, I believe the latter so strongly, if I can’t convince the others in the raft with me, I’m jumping in and swimming to another raft after my entire adult life here. 

Mark

On Sun, Nov 3, 2019 at 7:30 AM Mark Miller <[hidden email]> wrote:
In fact this will be a fundamental difference some of us are about to split between.

Those that think they can ever fix the tests or the system or the 1000s of bugs we have and keep adding due to our current world view of making tests fit the system not the system fit the tests and that fact that everything is so slow and retry and workaround that stupid shit works all over. It's all deep. It's ingrained. It grown over for a decade.Its a project of 60 modules.

Soon we will split between those that think they are making progress across the ocean and those that think we are sitting in shark infested waters waiting to die actually, starting to float backwards sometimes now.

- Mark

On Sun, Nov 3, 2019 at 7:23 AM Mark Miller <[hidden email]> wrote:
bq.  They also would allow it to do it in an iterative manner without changing everything at once.

Sadly, you can't fix this piece by piece :) I dare anyone to try. I encourage, I applaud the effort.

The world is your oyster from a good spot - take your pick of how to do things.

But from this spot, if anyone thinks we are getting out design change by design change, JIRA by JIRA, I'm so sorry. Let's commiserate in a couple years on a beer when you  give up on that.

- Mark

On Sun, Nov 3, 2019 at 4:01 AM Jörn Franke <[hidden email]> wrote:
I cannot say anything about the statements, but maybe it could help to introduce Solr Improvement Proposals (SIP) similar to Kafka Improvement Proposals (KIP) or Flink Improvement Proposals (FLIP).

  I think they are helpful to facilitate design decisions and refactoring / redesign decision. They also would allow it to do it in an iterative manner without changing everything at once.
The final version could be out  in The Git of Solr in markdown including figures presenting parts of the design.

However for developing them I propose a more inclusive approach where many people (not only core developers) can easily comment and support, eg Google docs or similar.

> Am 03.11.2019 um 06:39 schrieb Noble Paul <[hidden email]>:
>
> Solr has to do more than Lucene. A Lucene user is mostly a developer
> who reads javadocs. A Solr user's touch points are
>
> * Public API
> * Ref guide
> * publicly visible files (in ZK as well as file system)
> * What to see/look for in the log files to debug issues
>
> Then we have more nuanced touch points such as the knowledge base of
> what happens internally in the system when 'X' API is invoked or when
> 'Y' behavior is observed in ZK data.
>
> The problem with delaying the review process till code completion is
> that, any changes based on review comments will require massive amount
> of work.
>
> I don't have an answer to how we achieve it. But, I clearly see this
> as a major gap in our development process today.
>
> This discussion may not be relevant in this thread, may be because no
> behavior is changed at all. We don't know yet
>
> What I want to believe is Mark is doing the right thing & it's gonna
> help us all in dealing with our operational issues. I don't want to
> interrupt his work with more discussions.
>
> Thanks you
>
>
>> On Sun, Nov 3, 2019 at 3:32 PM David Smiley <[hidden email]> wrote:
>>
>> Yeah we do a bad job of the things you listed Noble.  :-(   My colleagues want pointers to internal docs but the sad reality is there isn't any.  You may notice I'm a stickler in my code reviews for requiring javadocs on all top level classes.  I think more javadocs and code comments would be very helpful -- especially for the major classes.  This might help us all and others a lot more.  For example I think Lucene does a rather fine job of this for its major classes -- IndexWriter being a good example.
>>
>> ~ David Smiley
>> Apache Lucene/Solr Search Developer
>> http://www.linkedin.com/in/davidwsmiley
>>
>>
>>> On Sat, Nov 2, 2019 at 7:32 PM Noble Paul <[hidden email]> wrote:
>>>
>>> Hi,
>>>
>>> I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:
>>>
>>> * How does it work?
>>>
>>> * What are the performance characteristics? Can it be done more efficiently?
>>>
>>> * What are the public touch points?
>>>
>>> ** Which are the files we store in ZK? Are they expected to be watched always?
>>>
>>> ** Or are they read on demand?
>>>
>>> ** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?
>>>
>>>
>>> We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt.
>>>
>>>
>>> This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.)
>>>
>>>
>>> Now, why am I discussing this in this thread?
>>>
>>>
>>> While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.
>>>
>>>
>>> I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion.
>>>
>>>
>>> I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
>>> Thanks
>>>
>>> On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
>>>>
>>>> Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).
>>>>
>>>> What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?
>>>>
>>>> On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
>>>>>
>>>>> Give me a short bit to follow up and I will lay out my case and proposal.
>>>>>
>>>>> Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.
>>>>>
>>>>> If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.
>>>>>
>>>>> - mark
>>>>>
>>>>> On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
>>>>>>
>>>>>> Hey All Solr Dev's,
>>>>>>
>>>>>> SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).
>>>>>>
>>>>>> Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.
>>>>>>
>>>>>> This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.
>>>>>>
>>>>>> There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.
>>>>>>
>>>>>> So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.
>>>>>>
>>>>>> We should all be embarrassed by the state of things.
>>>>>>
>>>>>> I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.
>>>>>>
>>>>>> --
>>>>>> - Mark
>>>>>>
>>>>>> http://about.me/markrmiller
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> - Mark
>>>>>
>>>>> http://about.me/markrmiller
>
>
>
> --
> -----------------------------------------------------
> Noble Paul
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: [hidden email]
> For additional commands, e-mail: [hidden email]
>

---------------------------------------------------------------------
To unsubscribe, e-mail: [hidden email]
For additional commands, e-mail: [hidden email]



--


--
--


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Mark Miller-3
And the bummer is, in the midst of this madness people are doing good work. Good cleanups. Good improvements. Good features. Good code. And it’s all basically wasted. It’s my hurts my mind. 

Mark

On Sun, Nov 3, 2019 at 7:58 AM Mark Miller <[hidden email]> wrote:
From a credentials standpoint:

Yonik and I built 90% of it originally and then I spent years on it with few other devs or users.

Pretty sure I'm the only one that has ever had 95%+ of the Solr test suites work in under 10-15 seconds consistently - 4000 tests across like 1000 suites. Got them all to run in parallel in under 5 minutes vs the 20-45 it takes on a good day after tons of other speed ups and fixes I've already done.

I doubt there are many insane enough that have deep dived and pushed around the entire code base for 2-3 weeks multiple times, 16-20 hours a day. That has spent the last decade beyond that stupid time almost exclusively on this system. Designing it with Yonik, building it, fixing it, helping people with it, monitoring it, responding to pages and escalations for it.

That has spent half a year replacing the entire decade old build for Lucene and Solr and all it's various nooks and crannies.

If anyone has spent more time on this system or pushing it around on a large scale or or has seen it in 100x the shape it is now more than once, please speak up, you are in charge, I follow you. I'm not that bright, if you have done the ground work, default to you.

Otherwise, I don't even have much confidence anyone else even knows this system remotely well. All that time and effort and the most I know of it is what awful awful shape its in and the bad trend direction.


- Mark

On Sun, Nov 3, 2019 at 7:35 AM Mark Miller <[hidden email]> wrote:
Personally, I believe the latter so strongly, if I can’t convince the others in the raft with me, I’m jumping in and swimming to another raft after my entire adult life here. 

Mark

On Sun, Nov 3, 2019 at 7:30 AM Mark Miller <[hidden email]> wrote:
In fact this will be a fundamental difference some of us are about to split between.

Those that think they can ever fix the tests or the system or the 1000s of bugs we have and keep adding due to our current world view of making tests fit the system not the system fit the tests and that fact that everything is so slow and retry and workaround that stupid shit works all over. It's all deep. It's ingrained. It grown over for a decade.Its a project of 60 modules.

Soon we will split between those that think they are making progress across the ocean and those that think we are sitting in shark infested waters waiting to die actually, starting to float backwards sometimes now.

- Mark

On Sun, Nov 3, 2019 at 7:23 AM Mark Miller <[hidden email]> wrote:
bq.  They also would allow it to do it in an iterative manner without changing everything at once.

Sadly, you can't fix this piece by piece :) I dare anyone to try. I encourage, I applaud the effort.

The world is your oyster from a good spot - take your pick of how to do things.

But from this spot, if anyone thinks we are getting out design change by design change, JIRA by JIRA, I'm so sorry. Let's commiserate in a couple years on a beer when you  give up on that.

- Mark

On Sun, Nov 3, 2019 at 4:01 AM Jörn Franke <[hidden email]> wrote:
I cannot say anything about the statements, but maybe it could help to introduce Solr Improvement Proposals (SIP) similar to Kafka Improvement Proposals (KIP) or Flink Improvement Proposals (FLIP).

  I think they are helpful to facilitate design decisions and refactoring / redesign decision. They also would allow it to do it in an iterative manner without changing everything at once.
The final version could be out  in The Git of Solr in markdown including figures presenting parts of the design.

However for developing them I propose a more inclusive approach where many people (not only core developers) can easily comment and support, eg Google docs or similar.

> Am 03.11.2019 um 06:39 schrieb Noble Paul <[hidden email]>:
>
> Solr has to do more than Lucene. A Lucene user is mostly a developer
> who reads javadocs. A Solr user's touch points are
>
> * Public API
> * Ref guide
> * publicly visible files (in ZK as well as file system)
> * What to see/look for in the log files to debug issues
>
> Then we have more nuanced touch points such as the knowledge base of
> what happens internally in the system when 'X' API is invoked or when
> 'Y' behavior is observed in ZK data.
>
> The problem with delaying the review process till code completion is
> that, any changes based on review comments will require massive amount
> of work.
>
> I don't have an answer to how we achieve it. But, I clearly see this
> as a major gap in our development process today.
>
> This discussion may not be relevant in this thread, may be because no
> behavior is changed at all. We don't know yet
>
> What I want to believe is Mark is doing the right thing & it's gonna
> help us all in dealing with our operational issues. I don't want to
> interrupt his work with more discussions.
>
> Thanks you
>
>
>> On Sun, Nov 3, 2019 at 3:32 PM David Smiley <[hidden email]> wrote:
>>
>> Yeah we do a bad job of the things you listed Noble.  :-(   My colleagues want pointers to internal docs but the sad reality is there isn't any.  You may notice I'm a stickler in my code reviews for requiring javadocs on all top level classes.  I think more javadocs and code comments would be very helpful -- especially for the major classes.  This might help us all and others a lot more.  For example I think Lucene does a rather fine job of this for its major classes -- IndexWriter being a good example.
>>
>> ~ David Smiley
>> Apache Lucene/Solr Search Developer
>> http://www.linkedin.com/in/davidwsmiley
>>
>>
>>> On Sat, Nov 2, 2019 at 7:32 PM Noble Paul <[hidden email]> wrote:
>>>
>>> Hi,
>>>
>>> I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:
>>>
>>> * How does it work?
>>>
>>> * What are the performance characteristics? Can it be done more efficiently?
>>>
>>> * What are the public touch points?
>>>
>>> ** Which are the files we store in ZK? Are they expected to be watched always?
>>>
>>> ** Or are they read on demand?
>>>
>>> ** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?
>>>
>>>
>>> We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt.
>>>
>>>
>>> This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.)
>>>
>>>
>>> Now, why am I discussing this in this thread?
>>>
>>>
>>> While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.
>>>
>>>
>>> I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion.
>>>
>>>
>>> I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
>>> Thanks
>>>
>>> On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
>>>>
>>>> Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).
>>>>
>>>> What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?
>>>>
>>>> On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
>>>>>
>>>>> Give me a short bit to follow up and I will lay out my case and proposal.
>>>>>
>>>>> Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.
>>>>>
>>>>> If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.
>>>>>
>>>>> - mark
>>>>>
>>>>> On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
>>>>>>
>>>>>> Hey All Solr Dev's,
>>>>>>
>>>>>> SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).
>>>>>>
>>>>>> Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.
>>>>>>
>>>>>> This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.
>>>>>>
>>>>>> There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.
>>>>>>
>>>>>> So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.
>>>>>>
>>>>>> We should all be embarrassed by the state of things.
>>>>>>
>>>>>> I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.
>>>>>>
>>>>>> --
>>>>>> - Mark
>>>>>>
>>>>>> http://about.me/markrmiller
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> - Mark
>>>>>
>>>>> http://about.me/markrmiller
>
>
>
> --
> -----------------------------------------------------
> Noble Paul
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: [hidden email]
> For additional commands, e-mail: [hidden email]
>

---------------------------------------------------------------------
To unsubscribe, e-mail: [hidden email]
For additional commands, e-mail: [hidden email]



--


--
--


--
--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Martin Gainty
In reply to this post by david.w.smiley@gmail.com
here is a bug i cannot shake in when building lucne/site

inside lucene/src/main/xml/ENTITY_TermQuery.xml

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE TermQuery [
<!ENTITY internalTerm "sumitomo">
<!ENTITY externalTerm SYSTEM "http://www.bar.xyz/external">
<!ENTITY % myParameterEntity "http://www.bar.xyz/param">
....

using ant build.xml:
 <!--
      The XSL input file is ignored completely, but XSL expects one to be given,
      so we pass ourself (${ant.file}) here. The list of module build.xmls is given
      via string parameter, that must be splitted by the XSL at '|'.
    -->
    <xslt in="${ant.file}" out="${javadoc.dir}/index.html" style="site/xsl/index.xsl" force="true">
      <outputproperty name="method" value="html"/>
      <outputproperty name="version" value="4.0"/>
      <outputproperty name="encoding" value="UTF-8"/>
      <outputproperty name="indent" value="yes"/>
      <param name="buildfiles" expression="${process-webpages.buildfiles}"/>
      <param name="version" expression="${version}"/>
      <param name="defaultCodec" expression="${defaultCodec}"/>
    </xslt>

OR maven pom.xml 
  <plugin>
        <groupId>org.codehaus.mojo</groupId>
   <artifactId>xml-maven-plugin</artifactId>
           <version>1.0.1</version>
           <executions>
            <execution>
                 <id>validate</id>
                        <phase>initialize</phase>
                        <goals>
                         <goal>transform</goal>
                        </goals>
                        <configuration>
                           <forceCreation>true</forceCreation>
                           <skip>false</skip>
                           <outputDirectory>${project.build.directory}/target</outputDirectory>
     <transformationSets>
       <transformationSet>
  <dir>src/main/xml</dir>
  <stylesheet>C:/Maven-plugin/lucene-solr/lucene/site/xsl/index.xsl</stylesheet>
  <parameters>
   <parameter>
     <name>MyParam</name>
     <value>true</value>
   </parameter>
       </parameters>
       </transformationSet>
     </transformationSets>
           </configuration>
           </execution>
       </executions>
       <dependencies>
        <dependency>
         <groupId>net.sf.saxon</groupId>
         <artifactId>Saxon-HE</artifactId>
         <version>9.9.1-1</version>
        </dependency>
       </dependencies>
      </plugin>

either build executing XSLT i get the same error:

[ERROR] Failed to execute goal org.codehaus.mojo:xml-maven-plugin:1.0.1:transform (validate) on project analysis: Failed to transform input file lucene/src/main/xml/ENTITY_TermQuery.xml: I/O error reported by XML parser processing file://lucene/src/main/xml/ENTITY_TermQuery.xml: www.bar.xyz:
Unknown host www.bar.xyz
]>

apparently www.bar.xyz host is supposed to be a placeholder
but for the life of me I cannot see where www.bar.zyz placeholder is replaced by a valid URL

(i havent used DTD in at least 10 years and i am way out of my element when trying to resolve)
any suggestions?
martin

From: David Smiley <[hidden email]>
Sent: Sunday, November 3, 2019 12:32 AM
To: Solr/Lucene Dev <[hidden email]>
Cc: Mark Miller <[hidden email]>
Subject: Re: SolrCloud is sick.
 
Yeah we do a bad job of the things you listed Noble.  :-(   My colleagues want pointers to internal docs but the sad reality is there isn't any.  You may notice I'm a stickler in my code reviews for requiring javadocs on all top level classes.  I think more javadocs and code comments would be very helpful -- especially for the major classes.  This might help us all and others a lot more.  For example I think Lucene does a rather fine job of this for its major classes -- IndexWriter being a good example.

~ David Smiley
Apache Lucene/Solr Search Developer


On Sat, Nov 2, 2019 at 7:32 PM Noble Paul <[hidden email]> wrote:
Hi, 

I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:

* How does it work?

* What are the performance characteristics? Can it be done more efficiently?

* What are the public touch points?

** Which are the files we store in ZK? Are they expected to be watched always?

** Or are they read on demand?

** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?


We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt. 


This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.) 


Now, why am I discussing this in this thread?


While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.


I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion. 


I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
Thanks

On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).

What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?

On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Jan Høydahl / Cominvent
Martin, please don’t hijack a subject but start a new mail thread for a new topic.

Jan Høydahl

4. nov. 2019 kl. 01:00 skrev Martin Gainty <[hidden email]>:


here is a bug i cannot shake in when building lucne/site

inside lucene/src/main/xml/ENTITY_TermQuery.xml

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE TermQuery [
<!ENTITY internalTerm "sumitomo">
<!ENTITY externalTerm SYSTEM "http://www.bar.xyz/external">
<!ENTITY % myParameterEntity "http://www.bar.xyz/param">
....

using ant build.xml:
 <!--
      The XSL input file is ignored completely, but XSL expects one to be given,
      so we pass ourself (${ant.file}) here. The list of module build.xmls is given
      via string parameter, that must be splitted by the XSL at '|'.
    -->
    <xslt in="${ant.file}" out="${javadoc.dir}/index.html" style="site/xsl/index.xsl" force="true">
      <outputproperty name="method" value="html"/>
      <outputproperty name="version" value="4.0"/>
      <outputproperty name="encoding" value="UTF-8"/>
      <outputproperty name="indent" value="yes"/>
      <param name="buildfiles" expression="${process-webpages.buildfiles}"/>
      <param name="version" expression="${version}"/>
      <param name="defaultCodec" expression="${defaultCodec}"/>
    </xslt>

OR maven pom.xml 
  <plugin>
        <groupId>org.codehaus.mojo</groupId>
   <artifactId>xml-maven-plugin</artifactId>
           <version>1.0.1</version>
           <executions>
            <execution>
                 <id>validate</id>
                        <phase>initialize</phase>
                        <goals>
                         <goal>transform</goal>
                        </goals>
                        <configuration>
                           <forceCreation>true</forceCreation>
                           <skip>false</skip>
                           <outputDirectory>${project.build.directory}/target</outputDirectory>
     <transformationSets>
       <transformationSet>
  <dir>src/main/xml</dir>
  <stylesheet>C:/Maven-plugin/lucene-solr/lucene/site/xsl/index.xsl</stylesheet>
  <parameters>
   <parameter>
     <name>MyParam</name>
     <value>true</value>
   </parameter>
       </parameters>
       </transformationSet>
     </transformationSets>
           </configuration>
           </execution>
       </executions>
       <dependencies>
        <dependency>
         <groupId>net.sf.saxon</groupId>
         <artifactId>Saxon-HE</artifactId>
         <version>9.9.1-1</version>
        </dependency>
       </dependencies>
      </plugin>

either build executing XSLT i get the same error:

[ERROR] Failed to execute goal org.codehaus.mojo:xml-maven-plugin:1.0.1:transform (validate) on project analysis: Failed to transform input file lucene/src/main/xml/ENTITY_TermQuery.xml: I/O error reported by XML parser processing file://lucene/src/main/xml/ENTITY_TermQuery.xml: www.bar.xyz:
Unknown host www.bar.xyz
]>

apparently www.bar.xyz host is supposed to be a placeholder
but for the life of me I cannot see where www.bar.zyz placeholder is replaced by a valid URL

(i havent used DTD in at least 10 years and i am way out of my element when trying to resolve)
any suggestions?
martin

From: David Smiley <[hidden email]>
Sent: Sunday, November 3, 2019 12:32 AM
To: Solr/Lucene Dev <[hidden email]>
Cc: Mark Miller <[hidden email]>
Subject: Re: SolrCloud is sick.
 
Yeah we do a bad job of the things you listed Noble.  :-(   My colleagues want pointers to internal docs but the sad reality is there isn't any.  You may notice I'm a stickler in my code reviews for requiring javadocs on all top level classes.  I think more javadocs and code comments would be very helpful -- especially for the major classes.  This might help us all and others a lot more.  For example I think Lucene does a rather fine job of this for its major classes -- IndexWriter being a good example.

~ David Smiley
Apache Lucene/Solr Search Developer


On Sat, Nov 2, 2019 at 7:32 PM Noble Paul <[hidden email]> wrote:
Hi, 

I believe there is a consensus on what is wrong with the way we have built the cluster state and overseer. We need to focus a bit more on the design aspect. Design, according to me, has the following elements:

* How does it work?

* What are the performance characteristics? Can it be done more efficiently?

* What are the public touch points?

** Which are the files we store in ZK? Are they expected to be watched always?

** Or are they read on demand?

** The public APIs. Does it make sense to the user? Can it be further simplified? How does it compare to the other APIs in the system?


We, as a community, do a bad job in dealing with these. While we focus on internal things, these are not discussed before it is too late. We usually do coding, tests, code review (sometimes) and commit. This leads to huge technical debt. 


This is not to put blame on one person or a group of people. (I occasionally see people discussing design issues upfront, I just hope that is the norm.) 


Now, why am I discussing this in this thread?


While we agree there are problems, we are trying to solve the problem using the same process we used to create these problems. Again, I'm not questioning the intent or competence of anyone. Unless we set the process right, we are doomed to make the same mistakes again.


I whole heartedly endorse any effort to improve SolrCloud/overseer. At the same time I fail to see us leveraging the collective experience of our community through meaningful discussion. 


I hope we don't resort to personal attacks and use this as an opportunity to improve our processes.
Thanks

On Sun, Nov 3, 2019, 9:52 AM Scott Blum <[hidden email]> wrote:
Very much agreed.  I've been trying to figure out for a long time what is the point in having a replica DOWN state that has to be toggled (DOWN and then UP!) every time a node restarts.  Considering that we could just combine ACTIVE and `live_nodes` to understand whether a replica is available.  It's not even foolproof since kill -9 on a solr node won't mark all the replicas DOWN-- that doesn't happen until the node comes back up (perversely).

What would it take to get to a state where restarting a node would require a minimal amount of ZK work in most cases?

On Sat, Nov 2, 2019 at 5:44 PM Mark Miller <[hidden email]> wrote:
Give me a short bit to follow up and I will lay out my case and proposal.

Everyone is then free to decide that we need to do something drastic or that I'm wrong and we should just continue down the same road. If that's the case, a lot of your work will get a lot easier and less impeded by me and we will still all be happier. Win win.

If we can just not make drastic changes for a just a brief week or so window, I'll say what I have to say, you guys can judge and do whatever you'd please.

- mark

On Fri, Nov 1, 2019 at 7:46 PM Mark Miller <[hidden email]> wrote:
Hey All Solr Dev's,

SolrCloud is sick right now. The way low level Zookeeper is handeled, the Overseer, is mix and mess of proper exception handling and super slow startup and shutdown, adding new things all the time with no concern for performance or proper ordering (which is harder to tell than you think).

Our class dependency graph doesn't even work - we just force it. Sort of. If the whole system  doesn't block and choke it's way to a start slow enough, lots of things fail.

This thing coughs up, you toss stuff into the storm, a good chunk of time, what you want eventually come back without causing too much damage.

There are so many things are are off or just plain wrong and the list is growing and growing. No one is following this or if you are, please back me up. This thing will collapse under it's own wait.

So if you want to add yet another state format cluster state or some other optimization on this junk heap, you can expect me to push back.

We should all be embarrassed by the state of things.

I've got some ideas for addressing them that I'll share soon, but god, don't keep optimizing a turd in non backcompat Overseer loving ways. That Overseer is an atrocity.


--
Reply | Threaded
Open this post in threaded view
|

Re: SolrCloud is sick.

Bram Van Dam
In reply to this post by Mark Miller-3
> SolrCloud is sick right now. The way low level Zookeeper is handeled

On an unrelated project, I've stopped using "raw" ZK client access and
have switched to Curator. The API is a fair bit easier to work with, and
it results in less ugly code. I realize that this won't go very far in
resolving more fundamental issues, but it might be something that can
help improve the shape of the code.

 - Bram

---------------------------------------------------------------------
To unsubscribe, e-mail: [hidden email]
For additional commands, e-mail: [hidden email]

12