This PR is the first step to use seq_no to optimize indexing operations.
The idea is to track the max seq_no of either update or delete ops on a
primary, and transfer this information to replicas, and replicas use it
to optimize indexing plan for index operations (with assigned seq_no).
The max_seq_no_of_updates on primary is initialized once when a primary
finishes its local recovery or peer recovery in relocation or being
promoted. After that, the max_seq_no_of_updates is only advanced internally
inside an engine when processing update or delete operations.
Relates #33656
This commit reverts most of #33157 as it introduces another race
condition and breaks a common case of watcher, when the first watch is
added to the system and the index does not exist yet.
This means, that the index will be created, which triggers a reload, but
during this time the put watch operation that triggered this is not yet
indexed, so that both processes finish roughly add the same time and
should not overwrite each other but act complementary.
This commit reverts the logic of cleaning out the ticker engine watches
on start up, as this is done already when the execution is paused -
which also gets paused on the cluster state listener again, as we can be
sure here, that the watches index has not yet been created.
This also adds a new test, that starts a one node cluster and emulates
the case of a non existing watches index and a watch being added, which
should result in proper execution.
Closes#33320
* Setting SO_LINGER for open but not connected non-blocking sockets
throws on OSX
* Fixed by only applying setting to connected sockets which will save
the same number of FDs as doing it on open sockets anyway
* closes#33879
Today we don't store the auto-generated timestamp of append-only
operations in Lucene; and assign -1 to every index operations
constructed from LuceneChangesSnapshot. This looks innocent but it
generates duplicate documents on a replica if a retry append-only
arrives first via peer-recovery; then an original append-only arrives
via replication. Since the retry append-only (delivered via recovery)
does not have timestamp, the replica will happily optimizes the original
request while it should not.
This change transmits the max auto-generated timestamp from the primary
to replicas before translog phase in peer recovery. This timestamp will
prevent replicas from optimizing append-only requests if retry
counterparts have been processed.
Relates #33656
Relates #33222
Currently, assertSeqNos assumes that the cluster is stable at the end of
the test (i.e., no more shard movement). However, this assumption does
not always hold. In these cases, we can stop the assertion instead of
failing a test.
Closes#33704
Changes the default of the `node.name` setting to the hostname of the
machine on which Elasticsearch is running. Previously it was the first 8
characters of the node id. This had the advantage of producing a unique
name even when the node name isn't configured but the disadvantage of
being unrecognizable and not being available until fairly late in the
startup process. Of particular interest is that it isn't available until
after logging is configured. This forces us to use a volatile read
whenever we add the node name to the log.
Using the hostname is available immediately on startup and is generally
recognizable but has the disadvantage of not being unique when run on
machines that don't set their hostname or when multiple elasticsearch
processes are run on the same host. I believe that, taken together, it
is better to default to the hostname.
1. Running multiple copies of Elasticsearch on the same node is a fairly
advanced feature. We do it all the as part of the elasticsearch build
for testing but we make sure to set the node name then.
2. That the node.name defaults to some flavor of "localhost" on an
unconfigured box feels like it isn't going to come up too much in
production. I expect most production deployments to at least set the
hostname.
As a bonus, production deployments need no longer set the node name in
most cases. At least in my experience most folks set it to the hostname
anyway.
Wraps all lines in our test framework at 140 characters because that is
our standard line length and removes all of the checkstyle suppressions
for the test framework.
Drops most of `ModuleTestCase` because it isn't used and we're moving
away from using guice in the way that it wants to test anyway. Also
switches a few classes that extend it but don't use it to extend
`ESTestCase` instead.
* TESTS: Set SO_LINGER = 0 for MockNioTransport
* Prevents lingering sockets in TIME_WAIT piling up during test runs and leading to port collisions that manifest as timeouts
* Fixes#32552
This change fixes a bug introduced in 6.3 that prevents fields with an explicit
similarity to be updated. It also adds a test that checks this case for similarities
but also for analyzers since they could suffer from the same problem.
Closes#33611
Today we use a special unicast hosts provider, the `MockUncasedHostsProvider`,
in many integration tests, to deal with the dynamic nature of the allocation of
ports to nodes. However #33241 allows us to use file-based discovery to achieve
the same goal, so the special test-only `MockUncasedHostsProvider` is no longer
required.
This change removes `MockUncasedHostProvider` and replaces it with file-based
discovery in tests based on `EsIntegTestCase`.
For correctness we need to verify whether the history uuid of the leader
index shards never changes while that index is being followed.
* The history UUIDs are recorded as custom index metadata in the follow index.
* The follow api validates whether the current history UUIDs of the leader
index shards are the same as the recorded history UUIDs.
If not the follow api fails.
* While a follow index is following a leader index; shard follow tasks
on each shard changes api call verify whether their current history uuid
is the same as the recorded history uuid.
Relates to #30086
Co-authored-by: Nhat Nguyen <nhat.nguyen@elastic.co>
This change adds a `_source` only snapshot repository that allows to wrap
any existing repository as a _backend_ to snapshot only the `_source` part
including live docs markers. Snapshots taken with the `source` repository
won't include any indices, doc-values or points. The snapshot will be reduced in size and
functionality such that it requires full re-indexing after it's successfully restored.
The restore process will copy the `_source` data locally starts a special shard and engine
to allow `match_all` scrolls and searches. Any other query, or get call will fail with and unsupported operation exception. The restored index is also marked as read-only.
This feature aims mainly for disaster recovery use-cases where snapshot size is
a concern or where time to restore is less of an issue.
**NOTE**: The snapshot produced by this repository is still a valid lucene index. This change doesn't allow for any longer retention policies which is out of scope for this change.
When a replica starts following a newly promoted primary, it may have
some operations which don't exist on the new primary. Thus we need to
throw those operations to align a replica with the new primary. This can
be done by first resetting an engine from the safe commit, then replaying
the local translog up to the global checkpoint.
Relates #32867
Today the full cluster restart tests do not preserve cluster settings on
restart. This is a mistake because it is not an accurate reflection of
reality, we do not expect users to clear cluster settings when they
perform a full cluster restart. This commit makes it so that all full
cluster restart tests preserve settings on upgrade.
* LeafCollector.setScorer() now takes a Scorable
* Scorers may not have null Weights
* IndexWriter.getFlushingBytes() reports how much memory is being used by IW threads writing to disk
In some cases we want to skip wiping cluster settings after a REST
test. For example, one use-case would be in the full cluster restart
tests where want to test cluster settings before and after a full
cluster restart. If we wipe the cluster settings before the restart,
then it would not be possible to assert on them after the restart.
In some cases we want to deprecate a setting, and then automatically
upgrade uses of that setting to a replacement setting. This commit adds
infrastructure for this so that we can upgrade settings when recovering
the cluster state, as well as when such settings are dynamically applied
on cluster update settings requests. This commit only focuses on cluster
settings, index settings can build on this infrastructure in a
follow-up.
This commit adds license checks for the auto-follow implementation. We
check the license on put auto-follow patterns, and then for every
coordination round we check that the local and remote clusters are
licensed for CCR. In the case of non-compliance, we skip coordination
yet continue to schedule follow-ups.
This commit ensures that we bootstrap a new history_uuid when force
allocating a stale primary. A stale primary should never be the source
of an operation-based recovery to another shard which exists before the
forced-allocation.
Closes#26712
Change the logging infrastructure to handle when the node name isn't
available in `elasticsearch.yml`. In that case the node name is not
available until long after logging is configured. The biggest change is
that the node name logging no longer fixed at pattern build time.
Instead it is read from a `SetOnce` on every print. If it is unset it is
printed as `unknown` so we have something that fits in the pattern.
On normal startup we don't log anything until the node name is available
so we never see the `unknown`s.
Instead of passing DirectoryService which causes yet another dependency
on Store we can just pass in a Directory since we will just call
`DirectoryService#newDirectory()` on it anyway.
This change collapses all metrics aggregations classes into a single package `org.elasticsearch.aggregations.metrics`.
It also restricts the visibility of some classes (aggregators and factories) that should not be used outside of the package.
Relates #22868
When we rollover and index we write the conditions of the rollover that
the old index met into the old index. Loading this index metadata
requires a working `NamedXContentRegistry` that has been populated with
parsers from the rollover infrastructure. We had a few loads that didn't
use a working `NamedXContentRegistry` and so would fail if they ever
encountered an index that had been rolled over. Here are the locations
of the loads and how I fixed them:
* IndexFolderUpgrader - removed entirely. It existed to support opening
indices made in Elasticsearch 2.x. Since we only need this change as far
back as 6.4.1 which will supports reading from indices created as far
back as 5.0.0 we should be good here.
* TransportNodesListGatewayStartedShards - wired the
`NamedXContentRegistry` into place.
* TransportNodesListShardStoreMetaData - wired the
`NamedXContentRegistry` into place.
* OldIndexUtils - removed entirely. It existed to support the zip based
index backwards compatibility tests which we've since replaced with code
that actually runs old versions of Elasticsearch.
In addition to fixing the actual problem I added full cluster restart
integration tests for rollover which would have caught this problem and
I added an extra assertion to IndexMetaData's deserialization code which
will trip if we try to deserialize and index's metadata without a fully
formed `NamedXContentRegistry`. It won't catch if use the *wrong*
`NamedXContentRegistry` but it is better than nothing.
Closes#33316
This commit allows us to use different TranslogRecoveryRunner when
recovering an engine from its local translog. This change is a
prerequisite for the commit-based rollback PR.
Relates #32867
The main benefit of the upgrade for users is the search optimization for top scored documents when the total hit count is not needed. However this optimization is not activated in this change, there is another issue opened to discuss how it should be integrated smoothly.
Some comments about the change:
* Tests that can produce negative scores have been adapted but we need to forbid them completely: #33309Closes#32899
This commit is related to #32517. It allows an "server_name"
attribute on a DiscoveryNode to be propagated to the server using
the TLS SNI extentsion. This functionality is only implemented for
the netty security transport.
Now that types are unique per mapping we can retrieve the document mapper
without referencing the type. This fixes an NPE when stored fields are disabled.
For 6x we'll need a different fix since mappings can still have multiple types.
Relates #32941
This commit introduces the formal notion of a private setting. This
enables us to register some settings that we had previously not
registered as fully-fledged settings to avoid them being exposed via
APIs such as the create index API. For example, we had hacks in the
codebase to allow index.version.created to be passed around inside of
settings objects, but was not registered as a setting so that if a user
tried to use the setting on any API then they would get an
exception. This prevented users from setting index.version.created on
index creation, or updating it via the index settings API. By
introducing private settings, we can continue to reject these attempts,
yet now we can represent these settings as actual settings. In this
change, we register index.version.created as an actual setting. We do
not cutover all settings that we had been treating as private in this
pull request, it is already quite large due to moving some tests around
to account for the fact that some tests need to be able to set the
index.version.created. This can be done in a follow-up change.
Drops `Settings` from some logging ctors now that they are no longer
needed. This should allow us to stop passing `Settings` around to quite
as many places.
This PR integrates Lucene soft-deletes(LUCENE-8200) into Elasticsearch.
Highlight works in this PR include:
- Replace hard-deletes by soft-deletes in InternalEngine
- Use _recovery_source if _source is disabled or modified (#31106)
- Soft-deletes retention policy based on the global checkpoint (#30335)
- Read operation history from Lucene instead of translog (#30120)
- Use Lucene history in peer-recovery (#30522)
Relates #30086Closes#29530
---
These works have been done by the whole team; however, these individuals
(lexical order) have significant contribution in coding and reviewing:
Co-authored-by: Adrien Grand <jpountz@gmail.com>
Co-authored-by: Boaz Leskes <b.leskes@gmail.com>
Co-authored-by: Jason Tedor <jason@tedor.me>
Co-authored-by: Martijn van Groningen <martijn.v.groningen@gmail.com>
Co-authored-by: Nhat Nguyen <nhat.nguyen@elastic.co>
Co-authored-by: Simon Willnauer <simonw@apache.org>
This PR integrates Lucene soft-deletes(LUCENE-8200) into Elasticsearch.
Highlight works in this PR include:
- Replace hard-deletes by soft-deletes in InternalEngine
- Use _recovery_source if _source is disabled or modified (#31106)
- Soft-deletes retention policy based on the global checkpoint (#30335)
- Read operation history from Lucene instead of translog (#30120)
- Use Lucene history in peer-recovery (#30522)
Relates #30086Closes#29530
---
These works have been done by the whole team; however, these individuals
(lexical order) have significant contribution in coding and reviewing:
Co-authored-by: Adrien Grand jpountz@gmail.com
Co-authored-by: Boaz Leskes b.leskes@gmail.com
Co-authored-by: Jason Tedor jason@tedor.me
Co-authored-by: Martijn van Groningen martijn.v.groningen@gmail.com
Co-authored-by: Nhat Nguyen nhat.nguyen@elastic.co
Co-authored-by: Simon Willnauer simonw@apache.org
Some AbstractDisruptionTestCase tests start failing since we enabled
assertSeqNos (in #33130). They fail because the assertSeqNos assertion
queries cluster stats while the cluster is disrupted or not formed yet.
This commit switches to use the cluster state and shard stats directly
from the test cluster.
Closes#33251
This commit checks that when we manually add a class to
the codebase map, that it does in-fact not exist on the classpath
in a jar. This will only be true if we are using the test framework
externally such as when a user develops a plugin.
Exclude classes meant for newer versions than what we are auditing against, those classes won't be found. There's no reason to exclude JDK classes from newer versions, with this PR, we will not extract them in the first place.
The new implementation is functional equivalent with the old, ant based one.
It parses task standard error to get the missing classes and violations in the same way.
I considered re-using ForbiddenApisCliTask but Gradle makes it hard to build inheritance with tasks that have task actions , since the order of the task actions can't be controlled.
This inheritance isn't dully desired either as the third party audit task is much more opinionated and we don't want to expose some of the configuration.
We could probably extract a common base class without any task actions, but probably more trouble than it's worth.
Closes#31715
We generate slightly different NoOps in InternalEngine and
TransportShardBulkAction for the same failure.
1. InternalEngine uses Exception#getFailure to generate a message
without the class name: newOp [NoOp{seqNo=1, primaryTerm=1,
reason='Contexts are mandatory in context enabled completion field
[suggest_context]'}].
2. TransportShardBulkAction uses Exception#toString to generate a
message with the class name: NoOp{seqNo=1, primaryTerm=1,
reason='java.lang.IllegalArgumentException: Contexts are mandatory in
context enabled completion field [suggest_context]'}.
If a write operation fails while a replica is recovering, that replica
will possibly receive two different NoOps: one from recovery and one
from replication. These two different NoOps will trip
TranslogWriter#assertNoSeqNumberConflict assertion.
This commit ensures that we generate the same Noop for the same failure.
Closes#32986
Nhat Nguyen
Vladimir Dolzhenko
lipsill
Alexander Reelsen
Armin Braun
Tim Vernum
Nik Everett
Or Bin
Alpar Torok
Jim Ferenczi
David Turner
Martijn van Groningen
Simon Willnauer
Jason Tedor
Alan Woodward
Tanguy Leroux
Tim Brooks
Matt Weber
Jonathan Little