[
https://issues.apache.org/jira/browse/SOLR-16497?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17627331#comment-17627331
]
Torsten Bøgh Köster commented on SOLR-16497:
--------------------------------------------
[~dsmiley] take your time, the change is quite huge :-/
A little background on our Solr setup: We run on AWS in a SolrCloud setup.
Inside AWS we have an Autoscaling Group for each Solr shard. The ASG scales out
at a given CPU usage percentage (currently 50%). We have a custom Solr cluster
manager that assigns replicas to new Solr instances joining the SolrCloud. The
ASG assigns a shard tag to each of its instances. This tag is picked up by the
cluster manager to place the correct shard on the instance.
Unfortunately the Solr Autoscaling framework never really worked for us so we
had to build the cluster manager ourselves. On the other hand it handles domain
specific taks like index creation & validation.
I'm happy to answer any questions :-)
> Allow finer grained locking in SolrCores
> ----------------------------------------
>
> Key: SOLR-16497
> URL: https://issues.apache.org/jira/browse/SOLR-16497
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: search
> Affects Versions: 9.0, 8.11.2
> Reporter: Torsten Bøgh Köster
> Assignee: David Smiley
> Priority: Major
> Attachments: solrcores_locking.png, solrcores_locking_fixed.png
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> Access to loaded SolrCore instances is a synchronized read and write
> operation in SolrCores#getCoreFromAnyList. This method is touched by every
> request as every HTTP request is assigned the SolrCore it operates on.
> h3. Background
> Under heavy load we discovered that application halts inside of Solr are
> becoming a serious problem in high traffic environments. Using Java Flight
> Recordings we discovered high accumulated applications halts on the
> modifyLock in SolrCores. In our case this means that we can only utilize our
> machines up to 25% cpu usage. With the fix applied, a utilization up to 80%
> is perfectly doable.
> In our case this specific locking problem was masked by another locking
> problem in the SlowCompositeReaderWrapper. We'll submit our fix to the
> locking problem in the SlowCompositeReaderWrapper in a following issue.
> h3. Problem
> Our Solr instances utilizes the collapse component heavily. The instances run
> with 32 cores and 32gb Java heap on a rather small index (4gb). The instances
> scale out at 50% cpu load. We take Java Flight Recorder snapshots of 60
> seconds
> as soon the cpu usage exceeds 50%.
> !solrcores_locking.png|width=1024!
> During our 60s Java Flight Recorder snapshot, the ~2k Jetty acceptor threads
> accumulated more than 12h locking time inside SolrCores on the modifyLock
> instance used as synchronized lock (see screenshot). With this fix the
> locking access is reduced to write accesses only. We validated this using
> another JFR snapshot:
> !solrcores_locking_fixed.png|width=1024!
> We ran this code for a couple of weeks in our live environment.
> h3. Solution
> The synchronized modifyLock is replaced by a ReentrantReadWriteLock. This
> allows concurrent reads from the internal SolrCore instance list (cores) but
> grants exclusive access to write operations on the instance list (cores).
> In Solr 9.x the cache inside the TransientSolrCoreCacheFactoryDefault adds a
> cache overflow handling of the size based internal cache (SOLR-15964). As
> soon as SolrCores are evicted from the internal cache, the cache behaviour
> changes from a size based cache to a reference based cache via the cache's
> eviction handler. SolrCore instances that are still referenced are inserted
> back into the cache. This means that write operations to the cache (insert
> SolrCore) can be issued during read operations in SolrCores. Hence these
> operations have only a read lock which cannot be upgraded to a write lock
> (dead lock).
> To overcome this, we moved the cache maintenance (including the eviction
> handler) in TransientSolrCoreCacheFactoryDefault to a separate thread. This
> thread can acquire a write lock but on the other hand a separate thread will
> schedule a ping-pong behaviour in the eviction handler on a full cache with
> SolrCores still referenced. To overcome this we made the overflow behaviour
> transparent by adding an additional overflowCores instance. Here we add
> evicted but still referenced cores from the transientCores cache.
> Furthermore we need to ensure that only a single transientSolrCoreCache
> inside TransientSolrCoreCacheFactoryDefault is created. As we now allow
> multiple read threads, we call the the getTransientCacheHandler() method
> initially holding a write lock inside the load() method. Calling the method
> only needs a write lock initially (for cache creation). For all other calls,
> a read lock is sufficient. By default, the getTransientCacheHandler()
> acquires a read lock. If a write is needed (e.g. for core creation), the
> getTransientCacheHandlerInternal() is called. This method explicitly does not
> use a lock in order to provide the flexibility to choose between a read-lock
> and a write-lock. This ensures that a single instance of
> transientSolrCoreCache is created.
> The lock signaling between SolrCore and CoreContainer gets replaced by a
> Condition that is tied to the write lock.
> h3. Summary
> This change allows for a finer grained access to the list of open SolrCores.
> The decreased blocking read access is noticeable in decreased blocking times
> of the Solr application (see screenshot).
> This change has been composed together by Dennis Berger, Torsten Bøgh Köster
> and Marco Petris.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
