Netty plays a lot of games with recycling byte buffers in thread local
caches, and using a pooled byte buffer allocator to reduce pressure on
the garbage collector.
The recycler in particular appears to be fraught with peril. It appears
that there are circumstances where the recycler does not recycle quickly
enough and can exceed its capacity leading to heap exhaustion and out of
memory errors. If you spend a few minutes reading the history of the
recycler on the Netty GitHub issues, it appears it has been nothing but
a source of trouble, and the project itself has an open issue that
proposes disabling by default and possibly even removing the recycler.
The pooled byte buffer allocator has problems itself. It sizes the pool
based on the number of runtime processors and can indeed grab a very
large percentage of the heap (in some cases 50% or more). Additionally,
the Netty project continues to struggle with leaks here.
We are seeing users struggle with issues in 5.x that I think are largely
driven by some of the problems here with Netty.
This change proposes to disable the recycler, and to disable the pooled
byte buffer allocator. I think that disabling these features will return
some of the stablity that these features appear to be losing us.
I have done performance testing on my workstation with disabling these
and I do not see a difference in performance. I propose that we make
this change in master and let some nightly benchmarks run to confirm
that there is not a difference in performance. If we are comfortable
with the performance changes, I propose backporting this to all active
branches.
Relates #22452
ParseFieldMatcher as well as ParseFieldMatcherSupplier will be soon removed, hence the ObjectParser's context doesn't need to be a ParseFieldMatcherSupplier anymore. That will allow to remove ParseFieldMatcherSupplier's implementations, little by little.
The test currently checks that the recovering shard is not failed when it is not a primary relocation that has moved past the finalization step.
Checking if it has moved past that step is done by intercepting the request between the replication source and the target and checking if it has seen
then WAIT_FOR_CLUSTERSTATE action as this is the next action that is called after finalization. This action can, however, occur only after the shard was
already failed, and thus trip the assertion. This commit changes the check to look out for the FINALIZE action, independently of whether it succeeded or not.
#22325 changed the recovery retry logic to use unique recovery ids. The change also introduced an issue, however, which made it possible for the shard store to be closed under CancellableThreads, triggering assertions in the node locking logic. This commit limits the use of CancellableThreads only to the part where we wait on the old recovery target to be closed.
The RestHighLevelClient class takes as as an argument a low level client instance RestClient. The first method added is ping, which returns true if the call to HEAD / went ok and false if an IOException was thrown. Any other exception gets bubbled up.
There are two kinds of tests, a unit test (RestHighLevelClientTests) that verifies the interaction between high level and low level client, and an integration test (MainActionIT) which relies on an externally started es cluster to send requests to.
Today we execute the low level handshake on the TCP layer in #connectToNode.
If #openConnection is used directly, which is truly expert, no handshake is executed
which allows connecting to nodes that are not necessarily compatible. This change
moves the handshake to #openConnection to prevent bypassing this logic.
This integrates the mocksocket jar with elasticsearch tests. Mocksocket wraps actions requiring SocketPermissions in doPrivilege blocks. This will eventually allow SocketPermissions to be assigned to the mocksocket jar opposed to the entire elasticsearch codebase.
Previously, we could run into a situation where attempting to delete an
index due to a cluster state update would cause an unhandled exception
to bubble up to the ClusterService and cause the cluster state applier
to fail. The result of this situation is that the cluster state never
gets updated on the ClusterService because the exception happens before
all cluster state appliers have completed and the ClusterService only
updates the cluster state once all cluster state appliers have
successfully completed.
All other methods on IndicesService properly handle all exceptions and
not just IOExceptions, but there were two instances with respect to
index deletion where only IOExceptions where handled by the
IndicesService. If any other exception occurred during these delete
operations, the exception would be bubbled up to the ClusterService,
causing the aforementioned issues.
This commit ensures all methods in IndicesService properly capture all
types of Exceptions, so that the ClusterService manages to update the
cluster state, even in the presence of shard creation/deletion failures.
Note that the lack of updating the cluster state in the presence of such
exceptions can have many unintended consequences, one of them being
the tripping of the assertion in IndicesClusterStateService#removeUnallocatedIndices
where the assumption is that if there is an IndexService to remove with
an unassigned shard, then the index must exist in the cluster state, but if
the cluster state was never updated due to the aforementioned exceptions,
then the cluster state will not have the index in question.
Currently `geo_point` and `geo_shape` field are treated as `text` field by the field stats API and we
try to extract the min/max values with MultiFields.getTerms.
This is ok in master because a `geo_point` field is always a Point field but it can cause problem in 5.x (and 2.x) because the legacy
`geo_point` are indexed as terms.
As a result the min and max are extracted and then printed in the FieldStats output using BytesRef.utf8ToString
which can throw an IndexOutOfBoundException since it's not valid UTF8 strings.
This change ensure that we never try to extract min/max information from a `geo_point` field.
It does not add a new type for geo points in the fieldstats API so we'll continue to use `text` for this kind of field.
This PR is targeted to master even though we could only commit this change to 5.x. I think it's cleaner to have it in master too before we make any decision on
https://github.com/elastic/elasticsearch/pull/21947.
Fixes#22384
Notifications for request tracing are invoked concurrently and can still
be in flight once a tracer is installed in the test. This can lead to side-effects
since the test relied on exact invocations. This commit adds action filtering
to the test tracer to only count invocations for the relevant actions.
Closes#22418
This commit cleans up the comments in IndexShard related to sequence numbers, making
them uniform in their formatting and taking advantage of the line-length
limit of 140 characters.
This commit cleans up the comments in GlobalCheckpointService, making
them uniform in their formatting and taking advantage of the line-length
limit of 140 characters.
This commit cleans up the comments in SequenceNumbersService, making
them uniform in their formatting and taking advantage of the line-length
limit of 140 characters.
Jason Tedor
Johannes Kanavin
Jason Veatch
javanna
Yannick Welsch
Adrien Grand
Nik Everett
Simon Willnauer
Ryan Ernst
Tim B
Ali Beyad
Jim Ferenczi
Simon Willnauer
Mike Dias