Currently, we treat all write operation exceptions as equals, but in reality
every write operation can cause either an environment failure (i.e. a failure
that should fail the engine e.g. data corruption, lucene tragic events) or
operation failure (i.e. a failure that is transient w.r.t the operation e.g.
parsing exception).
This change bubbles up enironment failures from the engine, after failing the
engine but captures transient operation failures as part of the operation
to be processed appopriately at the transport level.
Today we don't parse alias filters on the coordinating node, we only forward
the alias patters to executing node and resolve it late. This has several problems
like requests that go through filtered aliases are never cached if they use date math,
since the parsing happens very late in the process even without rewriting. It also used
to be processed on every shard while we can only do it once per index on the coordinating node.
Another nice side-effect is that we are never prone to cluster-state updates that change an alias,
all nodes will execute the exact same alias filter since they are process based on the same
cluster state.
Today Elasticsearch limits the number of processors used in computing
thread counts to 32. This was from a time when Elasticsearch created
more threads than it does now and users would run into out of memory
errors. It appears the real cause of these out of memory errors was not
well understood (it's often due to ulimit settings) and so users were
left hitting these out of memory errors on boxes with high core
counts. Today Elasticsearch creates less threads (but still a lot) and
we have a bootstrap check in place to ensure that the relevant ulimit is
not too low.
There are some caveats still to having too many concurrent indexing
threads as it can lead to too many little segments, and it's not a
magical go faster knob if indexing is already bottlenecked by disk, but
this limitation is artificial and surprising to users and so it should
be removed.
This commit also increases the lower bound of the max processes ulimit,
to prepare for a world where Elasticsearch instances might be running
with more the previous cap of 32 processors. With the current settings,
Elasticsearch wants to create roughly 576 + 25 * p / 2 threads, where p
is the number of processors. Add in roughly 7 * p / 8 threads for the GC
threads and a fudge factor, and 4096 should cover us pretty well up to
256 cores.
Relates #20874
Some people apparently never run tests when they change this file.
Neither do they read comments right below the line they change that
they should do the change after all.
It is important that folks understand that snapshot/restore isn't
for archiving. It is appropriate for backup and disaster recovery
but not for archival over long periods of time because of version
incompatibility.
Closes#20866
`AbstractSearchAsyncAction` has only been tested in integration tests.
The infrastructure is rather critical and should be tested on a unit-test
level. This change takes the first step.
This changes the CacheBuilder methods that are used to set expiration times to accept a
TimeValue instead of long. Accepting a long can lead to issues where the incorrect value is
passed in as the time unit is not clearly identified. By using TimeValue the caller no longer
needs to worry about the time unit used by the cache or builder.
Before this change the `MultiMatchQuery` called the field types
`termQuery()` with a null context. This is not correct so this change
fixes this so the `MultiMatchQuery` now uses the `ShardQueryContext` it
stores as a field.
Relates to https://github.com/elastic/elasticsearch/pull/20796#pullrequestreview-3606305
Both netty3 and netty4 http implementation printed the default
toString representation of PortRange if ports couldn't be bound.
This commit adds a better default toString method to PortRange and
uses the string representation for the error message in the http
implementations.
The test testDataFileCorruptionDuringRestore expects failures to happen when accessing snapshot data. It would sometimes
fail however as MockRepository (by default) only simulates 100 failures.
Sometimes it's useful / needed to use unreleased Version constants but we should not add those to the Version.java class for several reasons ie. BWC tests and assertions along those lines. Yet, it's not really obvious how to do that so I added some comments and a simple test for this.
There was an issue with using fuzziness parameter in multi_match query that has
been reported in #18710 and was fixed in Lucene 6.2 that is now used on master.
In order to verify that fix and close the original issue this PR adds the test
from that issue as an integration test.
today we might release a bytes array more than once if the send listener
throws an exception but already has released the array. Yet, this is already fixed
in the BytesArray class we use in production to ensure 3rd party users don't release
twice but our mocks still enforce it.
The snapshot restore state tracks information about shards being restored from a snapshot in the cluster state. For example it records if a shard has been successfully restored or if restoring it was not possible due to a corruption of the snapshot. Recording these events is usually based on changes to the shard routing table, i.e., when a shard is started after a successful restore or failed after an unsuccessful one. As of now, there were two communication channels to transmit recovery failure / success to update the routing table and the restore state. This lead to issues where a shard was failed but the restore state was not updated due to connection issues between data and master node. In some rare situations, this lead to an issue where the restore state could not be properly cleaned up anymore by the master, making it impossible to start new restore operations. The following change updates routing table and restore state in the same cluster state update so that both always stay in sync. It also eliminates the extra communication channel for restore operations and uses standard cluster state listener mechanism to update restore listener upon successful
completion of a snapshot.
Areek Zillur
Christoph Büscher
Nik Everett
Jun Ohtani
Jason Tedor
Simon Willnauer
Tanguy Leroux
Simon Willnauer
Clinton Gormley
kunal642
Jay Modi
Uli Fahrer
Colin Goodheart-Smithe
Thibaud BARDIN
Yannick Welsch
Robin Clarke
Boaz Leskes