With this commit we change the default behavior of
BulkProcessor from not backing off when getting
EsRejectedExecutionException to backing off exponentially.
This commit modifies the behavior after publication of a new cluster
state to only invoke the reroute logic once per batch of shard failures
rather than once per shard failure.
This commit modifies the internal representation of the JVM flag
UseCompressedOops to just be a String. This means we can just store the
value of the flag or "unknown" directly so that we do not have to engage
in shenanigans with three-valued logic around a boxed boolean.
Relates #15489
With this commit we introduce limited retries with a backoff logic to BulkProcessor
when a bulk request has been rejeced with an EsRejectedExecutionException.
Fixes#14620.
Currently MetaDataMappingService parses the mapping updates, reserializes it and
finally calls MapperService.merge with the serialized mapping. Given that mapping
serialization only writes differences from the default, this is a bit unfair to
parsers since they can't know whether some option has been explicitly set or not.
Furthermore this can cause bugs with metadata fields given that these fields use
existing field types as defaults.
This commit changes MetaDataMappingService to call MapperService.merge with the
original mapping update.
This fixes the `lenient` parameter to be `missingClasses`. I will remove this boolean and we can handle them via the normal whitelist.
It also adds a check for sheisty classes (jar hell with the jdk).
This is inspired by the lucene "sheisty" classes check, but it has false positives. This check is more evil, it validates every class file against the extension classloader as a resource, to see if it exists there. If so: jar hell.
This jar hell is a problem for several reasons:
1. causes insanely-hard-to-debug problems (like bugs in forbidden-apis)
2. hides problems (like internal api access)
3. the code you think is executing, is not really executing
4. security permissions are not what you think they are
5. brings in unnecessary dependencies
6. its jar hell
The more difficult problems are stuff like jython, where these classes are simply 'uberjared' directly in, so you cant just fix them by removing a bogus dependency. And there is a legit reason for them to do that, they want to support java 1.4.
We currently randomly add a set of mock plugins to integ tests.
Sometimes it is necessary to omit this mock plugins, but other times you
may just want to suppress a particular mock plugin. For example, if you
have your own transport, you want to omit the asserting local transport
mock, since they would both try to set the transport.type.
This commit adds a callback for a cluster state task executor that will
be invoked if the execution of a batch of cluster state update tasks
led to a new cluster state and that new cluster state was successfully
published.
Closes#15482
This commit adds to JvmInfo the status of whether or not compressed
ordinary object pointers are enabled. Additionally, logging of the max
heap size and the status of the compressed ordinary object pointers flag
are provided on startup.
Relates #13187, relates elastic/elasticsearch-definitive-guide#455
When a client sends a request to fail a shard to the master, the current
behavior is that the master will submit the cluster state update task
and then immediately send a successful response back to the client;
additionally, if there are any failures while processing the cluster
state update task to fail the shard, then the client will never be
notified of these failures.
This commit modifies the master behavior when handling requests to fail
a shard. In particular, the master will now wait until successful
publication of the cluster state update before notifying the request
client that the shard is marked as failed; additionally, the client is
now notified of any failures during the execution of the cluster state
update task.
Relates #14252
When creating a metadata mapper for a new type, we reuse an existing
configuration from an existing type (if any) in order to avoid introducing
conflicts. However this field type that is provided is considered as both an
initial configuration and the default configuration. So at serialization time,
we might only serialize the difference between the current configuration and
this default configuration, which might be different to what is actually
considered the default configuration.
This does not cause bugs today because metadata mappers usually override the
toXContent method and compare the current field type with Defaults.FIELD_TYPE
instead of defaultFieldType() but I would still like to do this change to
avoid future bugs.
Today we only test this when writing sequentially. Yet, in practice we mainly
write concurrently, this commit adds a test that tests that concurrent writes with
sudden fatal failure will not corrupt our translog.
Relates to #15420
Before we only evaluated segments that yielded matches in parent aggs, which caused us to miss to evaluate child docs in segments we didn't have parent matches for.
The fix for this is stop remember in what segments we have matches for and simply evaluate all segments. This makes the code simpler and we can still quickly see if a segment doesn't hold child docs like we did before.
This commit is a trivial reorganization of
o/e/c/a/s/ShardStateAction.java. The primary motive is have all of the
shard failure handling grouped together, and all of the shard started
handling grouped together.
The `path` option allowed to index/store a field `a.b.c` under just `c` when
set to `just_name`. This "feature" has been removed in 2.0 in favor of `copy_to`
so we can remove the back compat in 3.x.
There are two ways that a field can be defined twice:
- by reusing the name of a meta mapper in the root object (`_id`, `_routing`,
etc.)
- by defining a sub-field both explicitly in the mapping and through the code
in a field mapper (like ExternalMapper does)
This commit adds new checks in order to make sure this never happens.
Close#15057
Today mappings are mutable because of two APIs:
- Mapper.merge, which expects changes to be performed in-place
- IncludeInAll, which allows to change whether values should be put in the
`_all` field in place.
This commit changes both APIs to return a modified copy instead of modifying in
place so that mappings can be immutable. For now, only the type-level object is
immutable, but in the future we can imagine making them immutable at the
index-level so that mapping updates could be completely atomic at the index
level.
Close#9365
This change adds back the http.type setting. It also cleans up all the
transport related guice code to be consolidated within the
NetworkModule (as transport and http related stuff is what and how ES
exposes over the network). The setter methods previously used by some
plugins to override eg the TransportService or HttpServerTransport are
removed, and those plugins should now register a custom implementation
of the class with a name and set that using the appropriate config
setting. Note that I think ActionModule should also be moved into here,
to sit along side the rest actions, but I left that for a followup.
closes#14148
This commit addresses two type inference issues that the IntelliJ source
editor struggles with when registering query builder prototypes in
o/e/i/q/IndicesQueriesRegistry.java and
o/e/i/q/f/ScoreFunctionParserMapper.java.
This commit adds explicit logging at the DEBUG level for cluster state
update failures. Currently this responsibility is left to the cluster
state task listener, but we should expliclty log these with a generic
message to address cases where the listener might not.
Relates #14899, relates #15016, relates #15023
This commit changes the behavior of the logging in
TransportBroadcastByNodeAction#onNodeFailure to only trace log
exceptions that are considered shard-not-available exceptions. This
makes the logging consistent with how these exceptions are handled in
the response.
Relates #14927
Migrated from ES-Hadoop. Contains several improvements regarding:
* Security
Takes advantage of the pluggable security in ES 2.2 and uses that in order
to grant the necessary permissions to the Hadoop libs. It relies on a
dedicated DomainCombiner to grant permissions only when needed only to the
libraries installed in the plugin folder
Add security checks for SpecialPermission/scripting and provides out of
the box permissions for the latest Hadoop 1.x (1.2.1) and 2.x (2.7.1)
* Testing
Uses a customized Local FS to perform actual integration testing of the
Hadoop stack (and thus to make sure the proper permissions and ACC blocks
are in place) however without requiring extra permissions for testing.
If needed, a MiniDFS cluster is provided (though it requires extra
permissions to bind ports)
Provides a RestIT test
* Build system
Picks the build system used in ES (still Gradle)
After HighlightBuilder implements Writable now, we can remove
the temporary solution for transporting the highlight section in
SearchSourceBuilder from the coordinating node to the shard as
BytesReference and use HighlightBuilder instead.
The top-level highlighter has many options that can be overwritten per
field. Currently there is very similar code for this in two places.
This PR pulls out the parsing of the common parameters into
AbstractHighlighterBuilder for better reuse and to keep parsing of
common parameters more consistent.
Today we are super lenient (how could I missed that for f**k sake) with failing
/ closing the translog writer when we hit an exception. It's actually worse, we allow
to further write to it and don't care what has been already written to disk and what hasn't.
We keep the buffer in memory and try to write it again on the next operation.
When we hit a disk-full expcetion due to for instance a big merge we are likely adding document to the
translog but fail to write them to disk. Once the merge failed and freed up it's diskspace (note this is
a small window when concurrently indexing and failing the shard due to out of space exceptions) we will
allow in-flight operations to add to the translog and then once we fail the shard fsync it. These operations
are written to disk and fsynced which is fine but the previous buffer flush might have written some bytes
to disk which are not corrupting the translog. That wouldn't be an issue if we prevented the fsync.
Closes#15333
This change removes hardcoded ports from cluster formation. It passes
port 0 for http and transport, and then uses a special property to have
the node log the ports used for http and transport (just for tests).
This does not yet work for multi node tests. This brings us one step
closer to working with --parallel.
This commit improves the handling of ThreadLocal Random instance
allocation in o.e.c.Randomness.
- the seed per instance is no longer fixed
- a non-dangerous race to create the ThreadLocal instance has been
removed
- encapsulated all state into an static nested class for safe and lazy
instantiation
This commit adds the following:
* SpatialStrategy documentation to the geo-shape reference docs.
* Updates relation documentation to geo-shape-query reference docs.
* Updates GeoShapeFiledMapper to set points_only to true if TERM strategy is used (to be consistent with documentation)
This option allows to force the xcontent type to use to store the `_source`
document. The default is to use the same format as the input format.
This commit makes this option ignored for 2.x indices and rejected for 3.0
indices.
This commit removes and now forbids all uses of
Collections#shuffle(List) and Random#<init>() across the codebase. The
rationale for removing and forbidding these methods is to increase test
reproducibility. As these methods use non-reproducible seeds, production
code and tests that rely on these methods contribute to
non-reproducbility of tests.
Instead of Collections#shuffle(List) the method
Collections#shuffle(List, Random) can be used. All that is required then
is a reproducible source of randomness. Consequently, the utility class
Randomness has been added to assist in creating reproducible sources of
randomness.
Instead of Random#<init>(), Random#<init>(long) with a reproducible seed
or the aforementioned Randomess class can be used.
Closes#15287
In commit fafeb3a, we've refactored REST response handling logic
and returned HTTP status names instead of HTTP status codes for
bulk item responses. With this commit we restore the original
behavior.
Checked with @bleskes.
This method currently allows to write arbitrary bytes in an xcontent stream.
I changed it so that it can only write data to the same stream as the xcontent
(the bos parameter is removed) and that it yells at you if you try to write
raw bytes that can't be recognized as xcontent. Also the logic to copy the
structure instead of appending the bytes directly if the source and target
are of a different xcontent type have been moved to the low-level
XContentGenerator.
Tons of ancient "benchmarks" exist in elasticsearch. These are main
methods that do some kind of construction of ES classes and time various
things. The problem with these is they are not maintained, and not run.
Refactorings that touch anything that is common in these classes is very
painful. Going through these, almost all would simply not work in 2.x
without modifications (because they do not set path.home).
This change removes the entire benchmark package. If someone needs to
run a benchmark like this, they can look at history for examples if
necessary (although these examples are often not realistic and should
just start real elasticsearch processes in a shell script). Longer term,
we should make this easier to do by having the build support adding real
benchmarks which can be run in jenkins (so we know they actually run,
instead of doing refactorings with pure guesswork as to whether the
benchmark would run correctly).
we are not ready for this yet:
```
if (shardRouting.primary() && shardRouting.isRelocationTarget() == false) {
throw new IllegalIndexShardStateException(shardId, state, "shard is not a replica");
}
```
IndexResponse, DeleteResponse and UpdateResponse share some logic. This can be unified to a single DocWriteResponse base class. On top, some replication actions are now not about write operations anymore. This commit renames ActionWriteResponse to ReplicationResponse
Last some toXContent is moved from the Rest layer to the actual response classes, for more code re-sharing.
Closes#15334
The test configuration with seed A23029712A7EFB34 overwhelmed the pool which is invoked
in TransportService#sendLocalRequest().
With this commit we reduce the maximum number of concurrent requests from 10 to 7 and
add the failure message to the test output on the failing assertion for easier analysis.
As a replacement use ExistsQueryBuilder inside a mustNot() clause.
So instead of using `new ExistsQueryBuilder(name)` now use:
`new BoolQueryBuilder().mustNot(new ExistsQueryBuilder(name))`.
Closes#14112
This makes some minor improvements (does not fix all problems!)
It reorders unicast disco in elasticsearch.yml to be right after the network host,
for better locality.
It removes the warning (unreleased) about publish addresses, lets try to really discourage setting
that unless you need to (behind a proxy server). Most people should be fine with `network.host`
Finally it reorganizes the network docs page a bit:
We add a table of 4 "basic" settings at the very beginning:
* network.host
* discovery.zen.ping.unicast.hosts
* http.port
* transport.tcp.port
The first two being the most important, which addresses to bind and talk to, and the other two
being the port numbers.
The rest of the stuff I tried to simplify and reorder under "advanced" headers.
This is just a quick stab, I still think we need more effort into this thing, but we gotta start somewhere.
The NodeBuilder is currently used to construct a Node. However, this is
really just yet-another-builder that wraps around a Settings.Builder
witha couple convenience methods. But there are very few uses of these
convenience methods. This change removes NodeBuilder, in favor of just
using the Node constructor.
Follow up for #15340
We test that bind with wilcard IP + fixed IP it raises an exception
We test binding multiple IPs
(cherry picked from commit 2cc5bb7)
The goal of this method is to know whether the xcontent impl knows how to
differenciate floats from doubles or longs from ints or if it's just guessing.
However, all implementations return true (which is correct for yaml and json,
but cbor and smile should be able to differenciate). I first tried to implement
this method correctly but it raised many issues because eg. most impls write a
long as an integer when it is small enough. So I suggest that we remove this
method and just treat cbor and smile like yaml and json, which is already what
is happening today anyway.
The HighlightBuilder should be able to procude a SeachContextHighlight
object which contains the merged global and field options, also
contains objects that can only be created on the index shard (like
the actual lucene Query object used during highlighting).
This is done by the build() method of the HighlighBuilder. Also
adding tests that make sure the produced SearchContextHighlighter is
similar to the one we would get when parsing the xContent directly
with the current HighlightParseElement.
RecoverySource uses the RateLimiter under a cancelable thread. The SimpleRateLimiter used in throws ThreadInterruptedException on interruption. We should treat it as InterruptedException
Currently almost all our fields accept the `analyzer` and `term_vector` settings
although they only make sense on text fields. This commit forbids those settings
on all fields but `string` and `_all` for indices created on or after version
2.2.0.
throw exception if a copy_to is within a multi field
Copy to within multi field is ignored from 2.0 on, see #10802.
Instead of just ignoring it, we should throw an exception if this
is found in the mapping when a mapping is added. For already
existing indices we should at least log a warning.
We remove the copy_to in any case.
related to #14946
This commit adds the infrastructure to make settings that are updateable
resetable and changes the application of updates to be transactional. This means
setting updates are either applied or not. If the application failes all values are rejected.
This initial commit converts all dynamic cluster settings to make use of the new infrastructure.
All cluster level dynamic settings are not resettable to their defaults or to the node level settings.
The infrastructure also allows to list default values and descriptions which is not fully implemented yet.
Values can be reset using a list of key or simple regular expressions. This has only been implemented on the java
layer yet. For instance to reset all recovery settings to their defaults a user can just specify `indices.recovery.*`.
This commit also adds strict settings validation, if a setting is unknown or if a setting can not be applied the entire
settings update request will fail.
this ensures the codebase URL matches the permission grant (see matching toRealPath in Security.java)
in the case of symlinks or other shenanigans.
this is best effort, if we really want to support symlinks in any way, we need
e.g. qa or vagrant tests that configure a bunch of symlinks for things and ensure that in jenkins.
this should be easier to do with gradle, as we can just create a symlink'd home if we want
Today we only handle correctly if the `ExecutionCancelledException` comes from the
local execution. Yet, this can also come from remove and should be handled identically.
This commit restores the chunk size of 512kb lost in a previous but unreleased
refactoring. At the same time it removes the configurability of:
* `indices.recovery.file_chunk_size` - now fixed to 512kb
* `indices.recovery.translog_ops` - removed without replacement
* `indices.recovery.translog_size` - now fixed to 512kb
* `indices.recovery.compress` - file chunks are not compressed due to lucene's compression but translog operations are.
The compress option is gone entirely and compression is used where it makes sense. On sending files of the index
we don't compress as we rely on the lucene compression for stored fields etc.
Relates to #15161
This commit cherry picks some infrastructure changes from the `feature/seq_no` branch to make merging from master easier.
More explicitly, IndexShard current have prepareIndex and prepareDelete methods that are called both on the primary as the replica, giving it a different origin parameter. Instead, this commits creates two explicit prepare*OnPrimary and prepare*OnReplica methods. This has the extra added value of not expecting the caller to use an Engine enum.
Also, the commit adds some code reuse between TransportIndexAction and TransportDeleteAction and their TransportShardBulkAction counter parts.
Closes#15282
The tribe node creates one local client node for each cluster it
connects to. Refactorings in #13383 broke this so that each local client
node now tries to load the full elasticsearch.yml that the real tribe
node uses.
This change fixes the problem by adding a TribeClientNode which is a
subclass of Node. The Environment the node uses is now passed in (in
place of Settings), and the TribeClientNode simply does not use
InternalSettingsPreparer.prepareEnvironment.
The tests around tribe nodes are not great. The existing tests pass, but
I also manually tested by creating 2 local clusters, and configuring and
starting a tribe node. With this I was able to see in the logs the tribe
node connecting to each cluster.
closes#13383
I don't recall of this property of any of our field mappers and it's not in our
docs so I suspect it's very old. The removal of this property will not fail
version upgrades since none of the field mappers use it in toXContent.
This commit removes some unneeded null checks from
IndexingMemoryController that were left over from the work in #15251,
and simplifies the try-catch block in
IndexingMemoryController#updateShardBuffers.
For the search refactoring the HighlightBuilder needs a way to
create new instances by parsing xContent. For bwc this PR start
by moving over and slightly modifying the parsing from
HighlighterParseElement and keeps parsing for top level highlighter
and field options separate. Also adding tests for roundtrip
of random builder (rendering it to xContent and parsing it and
making sure the original builder properties are preserved)
Since 2.2 we run all scripts with minimal privileges, similar to applets in your browser.
The problem is, they have unrestricted access to other things they can muck with (ES, JDK, whatever).
So they can still easily do tons of bad things
This PR restricts what classes scripts can load via the classloader mechanism, to make life more difficult.
The "standard" list was populated from the old list used for the groovy sandbox: though
a few more were needed for tests to pass (java.lang.String, java.util.Iterator, nothing scary there).
Additionally, each scripting engine typically needs permissions to some runtime stuff.
That is the downside of this "good old classloader" approach, but I like the transparency and simplicity,
and I don't want to waste my time with any feature provided by the engine itself for this, I don't trust them.
This is not perfect and the engines are not perfect but you gotta start somewhere. For expert users that
need to tweak the permissions, we already support that via the standard java security configuration files, the
specification is simple, supports wildcards, etc (though we do not use them ourselves).
This commit simplifies shard inactive debug logging to only log when the
physical shard is marked as inactive. This eliminates duplicate logging
that existed in IndexShard#checkIdle and
IndexingMemoryController#checkIdle, and eliminates excessive logging
that was occurring when the shard was already inactive as a result of
the work in #15252.
Currently, when a user tries to install an old plugin (pre 2.x) on a 2.x
node, the error message is cryptic (just printing the file path that was
missing, when looking for the descriptor). This improves the message to
be more explicit that the descriptor is missing, and suggests the
problem might be the plugin was built before 2.0.
closes#15197
This commit addresses some issues that arose during the review of #14899
but were lost during squash while integrating into master.
- the number of test threads is dropped to at most eight
- a local variable is renamed for clarity
- task priorities are randomized
This commit fixes a test bug in
ClusterService#testClusterStateBatchedUpdates. In particular, in the
case that an executor did not receive a task assignment from the random
assignments, it would not have an entry in the map of executors to
counts of assigned tasks. The fix is to just check if each executor has
an entry in the counts map.
This commit modifies IndexingMemoryController to be stateless. Rather
than statefully tracking the indexing status of shards,
IndexingMemoryController can grab all available shards, check their idle
state, and then resize the buffers based on the number of and which
shards are not idle.
The driver for this change is a performance regression that can arise in
some scenarios after #13918. One scenario under which this performance
regression can arise is if an index is deleted and then created
again. Because IndexingMemoryController was previously statefully
tracking the state of shards via a map of ShardIds, the new shards with
the same ShardIds as previously existing shards would not be detected
and therefore their version maps would never be resized from the
defaults. This led to an explosion in the number of merges causing a
degradation in performance.
Closes#15225
Today we only check mapping compatibility when adding mappers to the
lookup structure. However, at this stage, the mapping has already been merged
partially, so we can leave mappings in a bad state. This commit removes the
compatibility check from Mapper.merge entirely and performs it _before_ we
call Mapper.merge.
One minor regression is that the exception messages don't group together errors
that come from MappedFieldType.checkCompatibility and Mapper.merge. Since we
run the former before the latter, Mapper.merge won't even have a chance to let
the user know about conflicts if conflicts were discovered by
MappedFieldType.checkCompatibility.
Close#15049
The `translated` flag makes LineStringBuilder stateful and gets set
to true under certain conditions when building a Shape or Geometry
from the ShapeBuilder. This makes building operations not be idempotent,
so calling build() more than once on a LineStringBuilder might change the
builder itself. This PR fixes this by replacing the instance variable by
a local `translated` flag that is only updated internally during the
building process and created again on any subsequent calls to build()
or buildGeometry().
Failures to merge a mapping can either come as a MergeMappingException if they
come from Mapper.merge or as an IllegalArgumentException if they come from
FieldTypeLookup.checkCompatibility. I think we should settle on one: this pull
request replaces all usage of MergeMappingException with
IllegalArgumentException.
The ttl could be specified as a time value only via the REST layer. That is now possible via java api too, either as a string or as a proper TimeValue. The internal format in IndexRequest becomes now TimeValue, which will then still converted to a long before storing the document.
Closes#15047
- Supports ImmutableOpenIntMap besides java.util.Map and ImmutableOpenMap
- Map keys can be any value (not only String)
- Map values do not have to implement Diffable interface. In that case custom value serializer needs to be provided.
Several settings have been deprecated or are replaced with new settings after refactorings
in version 1.x. This commit removes the support for these settings.
The settings are:
* `index.shard.recovery.translog_size`
* `index.shard.recovery.translog_ops`
* `index.shard.recovery.file_chunk_size`
* `index.shard.recovery.concurrent_streams`
* `index.shard.recovery.concurrent_small_file_streams`
* `indices.recovery.max_size_per_sec`
When not in debug mode, we currently only print the message of an
exception. However, this is not usually useful without knowing what the
exception type was. This change makes cli tools use toString() on the
exception so we get the type + message.
This commit wraps the trace logging statements in
TransportBroadcastByNodeAction in trace enabled checks to avoid
unnecessarily allocating objects.
The most egregious offenders were the two trace logging statements in
BroadcastByNodeTransportRequestHandler#onShardOperation. Aside from the
usual object allocations that occur when invoking ESLogger#trace (the
allocated object array for the varargs Object... parameter), these two
logging statements were invoking ShardRouting#shortSummary generating a
bunch of char arrays and Strings (from the StringBuilder, and so a bunch
of array copies as well). In a scenario where there are a lot of shards
and this method is being invoked frequently (e.g., constantly hitting
the _stats endpoint), these two unprotected trace logging statements
were generating a lot of unnecessary allocations.
This commit modifies the handling of shard started cluster state updates
to use the general cluster state batching mechanism. An advantage of
this approach is we now get correct per-listener notification on
failures.
Daniel Mitterdorfer
Yannick Welsch
Boaz Leskes
Simon Willnauer
Jason Tedor
Adrien Grand
Michael McCandless
Robert Muir
Robert Muir
Jack Conradson
Ryan Ernst
Britta Weber
javanna
Simon Willnauer
Britta Weber
Martijn van Groningen
soinlv
Jim Ferenczi
Costin Leau
Christoph Büscher
Nicholas Knize
Isabel Drost-Fromm
Jun Ohtani
Areek Zillur
David Pilato
Colin Goodheart-Smithe
Paulo Martins
Myll
Jim Ferenczi
Lee Hinman