Node Recovery Questions

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

Node Recovery Questions

sean mcevoy
Hi All,

A few questions on the procedure here to recover a failed node:
http://docs.basho.com/riak/kv/2.2.3/using/repair-recovery/failed-node/

We lost a production riak server when AWS decided to delete a node and we plan on doing this procedure to replace it with a newly built node. A practice run in our QA environment has brought up some questions.

- How can I tell when everything has synched up? I thought I could just monitor the handoffs but these completed within 5 minutes of comitting the cluster changes, the data directories continued to grow rapidly in size for at least an hour. I assume that this was data being synched to the new node but how can I tell when it has completed from the user level? Or is it left up to AAE to sync the data?

- The size of the bitcask directory on the 4 original nodes is ~10GB, on the new node the size of this directory climbed to 1GB within an hour but hasn't moved much in the 4 days since. I know bitcask entries still exist until the periodic compaction but can it be right that its hanging on to 90% the disk space its using for dead data?

- Not directly related to the recovery procedure, but while one node of a five-node cluster is down how is the extra load distributed within the cluster? It will still keep 3 copies of each entry, right? Are the copies that would have been on the missing node all stored on the next node in the ring, or distributed all around the cluster?

Thanks in advance,
//Sean.

_______________________________________________
riak-users mailing list
[hidden email]
http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
Reply | Threaded
Open this post in threaded view
|

RE: Node Recovery Questions

Martin Sumner
Sean,

Some partial answers to your questions.

I don't believe force-replace itself will sync anything up - it just reassigns ownership (hence handoff happens very quickly).

Read repair would synchronise a portion of the data.  So if 10% of you data is read regularly, this might explain some of what you see.

AAE should also repair your data.  But if nothing has happened for 4 days, then that doesn't seem to be the case.  It would be worth checking the aae-status page (http://docs.basho.com/riak/kv/2.2.3/using/admin/riak-admin/#aae-status) to confirm things are happening.

I don't know if there are any minimum levels of data before bitcask will perform compaction.  There's nothing obvious in the code that wouldn't be triggered way before 90%.  I don't know if it will merge on the active file (the one currently being written to), but that is 2GB max size (configured through bitcask.max_file_size).

When you say the size of the bitcask directory - is this the size shared across all vnodes on the node?  I guess if each vnode has a single file <2GB, and there are multiple vnodes - something unexpected might happen here?  If bitcask does indeed not merge the file active for writing.

In terms of distribution around the cluster, if you have an n_val of 3 you should normally expect to see a relatively even distribution of the data on failure (certainly not it all going to one).  Worst case scenario is that 3 nodes get all the load from that one failed node. 

When a vnode is inaccessible, 3 (assuming n=3) fallback vnodes are selected to handle the load for that 1 vnode (as that vnode would normally be in 3 preflists, and commonly a different node will be asked to start a vnode for each preflist).  


I will try and dig later into bitcask merge/compaction code, to see if I spot anything else.

Martin



_______________________________________________
riak-users mailing list
[hidden email]
http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
Reply | Threaded
Open this post in threaded view
|

Re: Node Recovery Questions

Martin Sumner
Based on a quick read of the code, compaction in bitcask is performed only on "readable" files, and the current active file for writing is excluded from that list.  With default settings, that active file can grow to 2GB.  So it is possible that if objects had been replaced/deleted many times within the active file, that space will not be recovered if all the replacements amount to < 2GB per vnode.  So at these small data sizes - you may get a relatively significant discrepancy between an old and recovered node in terms of disk space usage.

On 8 August 2018 at 17:37, Martin Sumner <[hidden email]> wrote:
Sean,

Some partial answers to your questions.

I don't believe force-replace itself will sync anything up - it just reassigns ownership (hence handoff happens very quickly).

Read repair would synchronise a portion of the data.  So if 10% of you data is read regularly, this might explain some of what you see.

AAE should also repair your data.  But if nothing has happened for 4 days, then that doesn't seem to be the case.  It would be worth checking the aae-status page (http://docs.basho.com/riak/kv/2.2.3/using/admin/riak-admin/#aae-status) to confirm things are happening.

I don't know if there are any minimum levels of data before bitcask will perform compaction.  There's nothing obvious in the code that wouldn't be triggered way before 90%.  I don't know if it will merge on the active file (the one currently being written to), but that is 2GB max size (configured through bitcask.max_file_size).

When you say the size of the bitcask directory - is this the size shared across all vnodes on the node?  I guess if each vnode has a single file <2GB, and there are multiple vnodes - something unexpected might happen here?  If bitcask does indeed not merge the file active for writing.

In terms of distribution around the cluster, if you have an n_val of 3 you should normally expect to see a relatively even distribution of the data on failure (certainly not it all going to one).  Worst case scenario is that 3 nodes get all the load from that one failed node. 

When a vnode is inaccessible, 3 (assuming n=3) fallback vnodes are selected to handle the load for that 1 vnode (as that vnode would normally be in 3 preflists, and commonly a different node will be asked to start a vnode for each preflist).  


I will try and dig later into bitcask merge/compaction code, to see if I spot anything else.

Martin




_______________________________________________
riak-users mailing list
[hidden email]
http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
Reply | Threaded
Open this post in threaded view
|

Re: Node Recovery Questions

sean mcevoy
Hi Martin,
Thanks for taking the time.
Yes, by "size of the bitcask directory" I mean I did a "du -h --max-depth=1 bitcask", so I think that would cover all the vnodes. We don't use any other backends.
Those answers are helpful, will get back to this in a few days and see what I can determine about where our data physically lies. Might have more questions then.
Cheers,
//Sean.

On Wed, Aug 8, 2018 at 6:05 PM, Martin Sumner <[hidden email]> wrote:
Based on a quick read of the code, compaction in bitcask is performed only on "readable" files, and the current active file for writing is excluded from that list.  With default settings, that active file can grow to 2GB.  So it is possible that if objects had been replaced/deleted many times within the active file, that space will not be recovered if all the replacements amount to < 2GB per vnode.  So at these small data sizes - you may get a relatively significant discrepancy between an old and recovered node in terms of disk space usage.

On 8 August 2018 at 17:37, Martin Sumner <[hidden email]> wrote:
Sean,

Some partial answers to your questions.

I don't believe force-replace itself will sync anything up - it just reassigns ownership (hence handoff happens very quickly).

Read repair would synchronise a portion of the data.  So if 10% of you data is read regularly, this might explain some of what you see.

AAE should also repair your data.  But if nothing has happened for 4 days, then that doesn't seem to be the case.  It would be worth checking the aae-status page (http://docs.basho.com/riak/kv/2.2.3/using/admin/riak-admin/#aae-status) to confirm things are happening.

I don't know if there are any minimum levels of data before bitcask will perform compaction.  There's nothing obvious in the code that wouldn't be triggered way before 90%.  I don't know if it will merge on the active file (the one currently being written to), but that is 2GB max size (configured through bitcask.max_file_size).

When you say the size of the bitcask directory - is this the size shared across all vnodes on the node?  I guess if each vnode has a single file <2GB, and there are multiple vnodes - something unexpected might happen here?  If bitcask does indeed not merge the file active for writing.

In terms of distribution around the cluster, if you have an n_val of 3 you should normally expect to see a relatively even distribution of the data on failure (certainly not it all going to one).  Worst case scenario is that 3 nodes get all the load from that one failed node. 

When a vnode is inaccessible, 3 (assuming n=3) fallback vnodes are selected to handle the load for that 1 vnode (as that vnode would normally be in 3 preflists, and commonly a different node will be asked to start a vnode for each preflist).  


I will try and dig later into bitcask merge/compaction code, to see if I spot anything else.

Martin





_______________________________________________
riak-users mailing list
[hidden email]
http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
Reply | Threaded
Open this post in threaded view
|

Re: Node Recovery Questions

sean mcevoy
Hi Martin, List,

Just an update to let ye know how things went and what we learned.

We did the force-replace procedure to bring the new node into the cluster in place of the old one. I attached to the riak erlang shell and with a little hacking was able to get all the bitcask handles and then do a bitcask:fold/3 to count keys. This showed that only a small percentage of all keys were present on the new node, even after the handoffs and transfers had completed.

Following the instructions at the bottom of this page: https://docs.basho.com/riak/kv/2.2.0/using/repair-recovery/repairs/
I attached to the erlang shell again and ran these commands (replacing the IP with our actual IP) to force repairs on all vnodes:

{ok, Ring} = riak_core_ring_manager:get_my_ring().
Partitions = [P || {P, '[hidden email]'} <- riak_core_ring:all_owners(Ring)].
[riak_kv_vnode:repair(P) || P <- Partitions].

The progress was most easily monitored with: riak-admin handoff summary
and once complete the new node had the expected number of keys.

Counting the keys is more than a bit hacky and occasionally caused a seg fault if there was background traffic, so I don't recommend it in general. But it did allow us to verify where the data was in our test env and then we could trust the procedure without counting keys in production. Monitoring the size of the bitcask directory is a lot lower resolution but it is at least safe, the results were similar in test & production so it was sufficient to verify the above procedure.

So in short, when replacing a node the force-replace procedure doesn't actually cause data to be synched to the new node. The above erlang shell commands do force a sync.

Thanks for the support!
//Sean.

On Thu, Aug 9, 2018 at 11:25 PM sean mcevoy <[hidden email]> wrote:
Hi Martin,
Thanks for taking the time.
Yes, by "size of the bitcask directory" I mean I did a "du -h --max-depth=1 bitcask", so I think that would cover all the vnodes. We don't use any other backends.
Those answers are helpful, will get back to this in a few days and see what I can determine about where our data physically lies. Might have more questions then.
Cheers,
//Sean.

On Wed, Aug 8, 2018 at 6:05 PM, Martin Sumner <[hidden email]> wrote:
Based on a quick read of the code, compaction in bitcask is performed only on "readable" files, and the current active file for writing is excluded from that list.  With default settings, that active file can grow to 2GB.  So it is possible that if objects had been replaced/deleted many times within the active file, that space will not be recovered if all the replacements amount to < 2GB per vnode.  So at these small data sizes - you may get a relatively significant discrepancy between an old and recovered node in terms of disk space usage.

On 8 August 2018 at 17:37, Martin Sumner <[hidden email]> wrote:
Sean,

Some partial answers to your questions.

I don't believe force-replace itself will sync anything up - it just reassigns ownership (hence handoff happens very quickly).

Read repair would synchronise a portion of the data.  So if 10% of you data is read regularly, this might explain some of what you see.

AAE should also repair your data.  But if nothing has happened for 4 days, then that doesn't seem to be the case.  It would be worth checking the aae-status page (http://docs.basho.com/riak/kv/2.2.3/using/admin/riak-admin/#aae-status) to confirm things are happening.

I don't know if there are any minimum levels of data before bitcask will perform compaction.  There's nothing obvious in the code that wouldn't be triggered way before 90%.  I don't know if it will merge on the active file (the one currently being written to), but that is 2GB max size (configured through bitcask.max_file_size).

When you say the size of the bitcask directory - is this the size shared across all vnodes on the node?  I guess if each vnode has a single file <2GB, and there are multiple vnodes - something unexpected might happen here?  If bitcask does indeed not merge the file active for writing.

In terms of distribution around the cluster, if you have an n_val of 3 you should normally expect to see a relatively even distribution of the data on failure (certainly not it all going to one).  Worst case scenario is that 3 nodes get all the load from that one failed node. 

When a vnode is inaccessible, 3 (assuming n=3) fallback vnodes are selected to handle the load for that 1 vnode (as that vnode would normally be in 3 preflists, and commonly a different node will be asked to start a vnode for each preflist).  


I will try and dig later into bitcask merge/compaction code, to see if I spot anything else.

Martin





_______________________________________________
riak-users mailing list
[hidden email]
http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com