Hi Erik,
that is very valuable info I missed.
Shouldn't that belong into an issue about rework at REBALANCELEADERS?
With your explanation the use of a queue makes sense and now I see some of
the logic behind.
- there is the leader and the firstWatcher
- if firstWatcher goes down or is inactive that one should positioned
at end of queue which automatically gives the baton to the next in queue
- preferredLeader has precedence if state=active
- there can be only one preferredLeader per shard
Any more rules?
From debugging I know there are a lot of actions taken behind the scene
and many things can go wrong. Parts included are ZkStateReader, Overseer,
ShardLeaderElectionContextBase, ShardLeaderElectionContext, ElectionContext,
LeaderElector, some Requests and a HttpSolrCall to REJOINLEADERELECTION.
An info I see when rebalance goes wrong is from ShardLeaderElectionContextBase:
"No version found for ephemeral leader parent node, won't remove previous leader
registration."
Regards, Bernd
Am 09.01.19 um 17:22 schrieb Erick Erickson:
Executive summary:
The central problem is "how can I insert an ephemeral node
in a specific place in a ZK queue". The code could be much,
much simpler if there were a reliable way to do just that. I haven't
looked at more recent ZKs to see if it's possible, I'd love it if
there were a better way.
On to details:
bq. wonder if we can discover the "inventor" of all this and ask him
how it should work
Yeah, I can contact that clown. That would be me ;)
The way leader election works is a ZK recipe where each
ephemeral node only watches the one in front of it. When a
node is deleted, the one watching it is notified.
So let's say we have nodes in this order: 1 watched by 2
watched by 3... In this case 1 is the leader. If 2 should
disappear, 3 gets notified and now watches 1. Nothing
else really happens.
But if the nodes are 1, 2, 3 and 1 disappears 2 gets notified
and says, in effect "I'm first in line so I will be leader". 3
doesn't get any notification.
bq. I wonder why the queue is not rotated until the new and preferred
leader is at front (position 0)
Because then you'd have N leader changes where N is the number
of nodes between the preferred leader and the actual leader in
the leader election queue at the start. That could result in 100s of
changes when only 1 is desired. The original case for this was
exactly that, there could be 100s of shards and several tens
to 100s of replicas.
Hmmmm, I suppose you wouldn't have that many leader changes if
you sent the ephemeral node in the _second_ position to the end
of the queue until the preferredLeader became the one in the second
position. You'd still have a watch fired for every requeueing though. I
haven't measured the cost there. That would also be an added
burden for the Overseer, which has been overwhelmed in the past.
I'm not against that solution, I don't have any real data to
evaluate.
bq. ....is possible to have more than one electionNode with the
same sequence number.
Yeah, and it causes complexity, but I don't have a good way around it. This is
related to your sorting question. ZK itself has a simple ordering, sequential
sequence numbers. Having two the same is the only way I could see (actually I
patterned it off some other code) to insert an ephemeral node second. What you
have then is two nodes "tied" for second by having the same sequence numbers.
Which one ZK thinks is second (and
thus which one becomes the leader if the zero'th ephemeral node disappears)
is based on sorting which includes the session ID, so there's code in there that
has to deal with sending the non-preferred node that's tied for second to the
end of the queue. That's the code that got changed during various
refactorings that I didn't take part in, and the code that's messed up.
bq. Wherever I see any java code to get the content from the queue it
is sorted. Where is the sense of this?
This is implied by the above, but explicitly so the Java code can see the queue
in the same order that ZK does and "do the right thing". In this case
assign the preferred leader's ephemeral node with the same sequence number
that the current second-in-line has and move the current second-in-line to the
end of the queue.
All that said, this code was written several years ago and I haven't looked at
whether there are better methods available now. The actions that are necessary
are:
1> ensure that the preferredLeader is the only node watching the leader in the
leader election queue
2> re-queue the leader at the end of the leader election queue. Since we'd be
sure the preferredLeader is watching the leader, this action would elect
the proper node as the leader.
Hmmm, note to self. It would help folks in the future if I, you know, documented
those two points in the code. Siiggghhh.
Last night I found what I _think_ is the problem I was having. Note that the
current code has three problems. I think I have fixes for all of them:
1> assigning the preferredLeader (or any SHARDUNIQUE property) does not
properly remove that property from other replicas in the shard if
present. So you may have multiple preferredLeaders in a shard.
2> the code to resolve tied sequence numbers had been changed
during some refactoring so the wrong node could be elected.
3> the response from the rebalanceleaders command isn't very useful, it's
on my plate to fix that. Partly it was not reporting useful
info, and partly your comment from the other day that it returns
without verifying the leadership has actually changed is well taken. At
present, it just changes the election queue and assumes that the
right thing happens. The test code was supposed to point out when
that assumption was incorrect, but you know the story there.
Currently, the code is pretty ugly in terms of all the junk I put in trying to
track this down, but when I clean it up I'll put up a patch. I added some
code to restart some of the jettys in the test (it's now "@Slow:") that
catches the restart case. Additionally, I changed the test to force
unique properties to be concentrated on a particular node then issue the
BALANCESHARDUNIQUE command to make sure that <1> above
doesn't happen.
Meanwhile, if there's an alternative approach that's simpler I'd be all
for it.
Best,
Erick
On Wed, Jan 9, 2019 at 1:32 AM Bernd Fehling
<bernd.fehl...@uni-bielefeld.de> wrote:
Yes, your findings are also very strange.
I wonder if we can discover the "inventor" of all this and ask him
how it should work or better how he originally wanted it to work.
Comments in the code (RebalanceLeaders.java) state that it is possible
to have more than one electionNode with the same sequence number.
Absolutely strange.
I wonder why the queue is not rotated until the new and preferred
leader is at front (position 0)?
But why is it a queue anyway?
Wherever I see any java code to get the content from the queue it
is sorted. Where is the sense of this?
Also, the elctionNodes have another attribute with name "ephemeral".
Where is that for and why is it not tested in TestRebalanceLeaders.java?
Regards, Bernd
Am 09.01.19 um 02:31 schrieb Erick Erickson:
It's weirder than that. In the current test on master, the
assumption is that the node recorded as leader in ZK
is actually the leader, see
TestRebalanceLeaders.checkZkLeadersAgree(). The theory
is that the identified leader node in ZK is actually the leader
after the rebalance command. But you're right, I don't see
an actual check that the collection's status agrees.
That aside, though, there are several problems I'm uncovering
1> BALANCESHARDUNIQUE can wind up with multiple
"preferredLeader" properties defined. Some time between
the original code and now someone refactored a bunch of
code and missed removing a unique property if it was
already assigned and being assigned to another replica
in the same slice.
2> to make it much worse, I've rewritten the tests
extensively and I can beast the rewritten tests 1,000
times and no failures. If I test manually by just issuing
the commands, everything works fine. By "testing manually"
I mean (working with 4 Vms, 10 shards 4 replicas)
create the collection
issue the BALANCESHARDUNIQUE command
issue the REBALANCELEADERS command
However, if instead I
create the collection
issue the BALANCESHARDUNIQUE command
shut down 3 of 4 Solr instances so all the leaders
are on the same host.
restart the 3 instances
issue the REBALANCELEADERS command then
it doesn't work.
At least that's what I think I'm seeing, but it makes no
real sense yet.
So I'm first trying to understand why my manual test
fails so regularly, then I can incorporate that setup
into the unit test (I'm thinking of just shutting down
and restarting some of the Jetty instances).
But it's a total mystery to me why restarting Solr instances
should have any effect. But that's certainly not
something that happens in the current test so I have
hopes that tracking that down will lead to understanding
what the invalid assumption I'm making is and we can
test for that too.,
On Tue, Jan 8, 2019 at 1:42 AM Bernd Fehling
<bernd.fehl...@uni-bielefeld.de> wrote:
Hi Erick,
after some more hours of debugging the rough result is, who ever invented
this leader election did not check if an action returns the estimated
result. There are only checks for exceptions, true/false, new sequence
numbers and so on, but never if a leader election to the preferredleader
really took place.
If doing a rebalanceleaders to preferredleader I also have to check if:
- a rebalance took place
- the preferredleader has really become leader (and not anyone else)
Currently this is not checked and the call rebalanceleaders to preferredleader
is like a shot into the dark with hope of success. And thats why any
problems have never been discovered or reported.
Bernd
Am 21.12.18 um 18:00 schrieb Erick Erickson:
I looked at the test last night and it's...disturbing. It succeeds
100% of the time. Manual testing seems to fail very often.
Of course it was late and I was a bit cross-eyed, so maybe
I wasn't looking at the manual tests correctly. Or maybe the
test is buggy.
I beasted the test 100x last night and all of them succeeded.
This was with all NRT replicas.
Today I'm going to modify the test into a stand-alone program
to see if it's something in the test environment that causes
it to succeed. I've got to get this to fail as a unit test before I
have confidence in any fixes, and also confidence that things
like this will be caught going forward.
Erick
On Fri, Dec 21, 2018 at 3:59 AM Bernd Fehling
<bernd.fehl...@uni-bielefeld.de> wrote:
As far as I could see with debugger there is still a problem in requeing.
There is a watcher and it is recognized that the watcher is not a
preferredleader.
So it tries to locate a preferredleader with success.
It then calls makeReplicaFirstWatcher and gets a new sequence number for
the preferredleader replica. But now we have two replicas with the same
sequence number. One replica which already owns that sequence number and
the replica which got the new (and the same) number as new sequence number.
It now tries to solve this with queueNodesWithSameSequence.
Might be something in rejoinElection.
At least the call to rejoinElection seems right. For preferredleader it
is true for rejoinAtHead and for the other replica with same sequence number
it is false for rejoinAtHead.
A test case should have 3 shards with 3 cores per shard and should try to
set preferredleader to different replicas at random. And then try to
rebalance and check the results.
So far, regards, Bernd
Am 21.12.18 um 07:11 schrieb Erick Erickson:
I'm reworking the test case, so hold off on doing that. If you want to
raise a JIRA, though. please do and attach your patch...
On Thu, Dec 20, 2018 at 10:53 AM Erick Erickson <erickerick...@gmail.com> wrote:
Nothing that I know of was _intentionally_ changed with this between
6x and 7x. That said, nothing that I know of was done to verify that
TLOG and PULL replicas (added in 7x) were handled correctly. There's a
test "TestRebalanceLeaders" for this functionality that has run since
the feature was put in, but it has _not_ been modified to create TLOG
and PULL replicas and test with those.
For this patch to be complete, we should either extend that test or
make another that fails without this patch and succeeds with it.
I'd probably recommend modifying TestRebalanceLeaders to randomly
create TLOG and (maybe) PULL replicas so we'd keep covering the
various cases.
Best,
Erick
On Thu, Dec 20, 2018 at 8:06 AM Bernd Fehling
<bernd.fehl...@uni-bielefeld.de> wrote:
Hi Vadim,
I just tried it with 6.6.5.
In my test cloud with 5 shards, 5 nodes, 3 cores per node it missed
one shard to become leader. But noticed that one shard already was
leader. No errors or exceptions in logs.
May be I should enable debug logging and try again to see all logging
messages from the patch.
Might be they also changed other parts between 6.6.5 and 7.6.0 so that
it works for you.
I also just changed from zookeeper 3.4.10 to 3.4.13 which works fine,
even with 3.4.10 dataDir. No errors no complains. Seems to be compatible.
Regards, Bernd
Am 20.12.18 um 12:31 schrieb Vadim Ivanov:
Yes! It works!
I have tested RebalanceLeaders today with the patch provided by Endika Posadas.
(http://lucene.472066.n3.nabble.com/Rebalance-Leaders-Leader-node-deleted-when-rebalancing-leaders-td4417040.html)
And at last it works as expected on my collection with 5 nodes and about 400
shards.
Original patch was slightly incompatible with 7.6.0
I hope this patch will help to try this feature with 7.6
https://drive.google.com/file/d/19z_MPjxItGyghTjXr6zTCVsiSJg1tN20
RebalanceLeaders was not very useful feature before 7.0 (as all replicas were
NRT)
But new replica types made it very helpful to keep big clusters in order...
I wonder, why there is no any jira about this case (or maybe I missed it)?
Anyone who cares, please, help to create jira and improve this feature in the
nearest releaase
