Several field mappers have a null_value parameter, that allows you to specify a placeholder
value to insert into a document if the incoming value for that field is null. The default value
for this is always null, meaning "add no placeholder". However, we explicitly bar users from
setting this parameter directly to null (done in #7978, in order to fix an NPE).
This exclusion means that if a mapper is serialized with include_defaults, then we either need
to special-case null_value to ensure that it is not output when it holds the default value, or
we find that the resulting serialized form cannot be used to create a mapping. This stops us
doing some useful generic testing of mappers.
This commit permits null as a parameter value for null_value, and changes the tests to check
that it is a) permissible and b) applied without throwing errors. As part of the testing changes,
a new base class MapperServiceTestCase is refactored from MapperTestCase, holding
the various helper methods related to building mappings but not the single-mapper specific
abstract methods.
Closes#58823
Fixes wrong NaN comparison in error message generator in GeoPolygonDecomposer and PolygonBuilder.
Supersedes #48207
Co-authored-by: Pedro Luiz Cabral Salomon Prado <pedroprado010@users.noreply.github.com>
The recursive data.path FilePermission check is an extremely hot
codepath in Elasticsearch. Unfortunately the FilePermission check in
Java is extremely allocation heavy. As it iterates through different
file permissions, it allocates byte arrays for each Path component that
must be compared. This PR improves the situation by adding the recursive
data.path FilePermission it its own PermissionsCollection object which
is checked first.
The change #57936 introduced a dedicated thread pool for reads in system indices.
It also introduced a potential NPE in the case the index to read in not yet present in
the cluster state. This commit fixes that bug by using the getIndexSafe() instead of
just index() method when retrieving the index's metadata so that an INFE is thrown
if the index does not exist.
We had a bug here were we put a `null` value into the shard
assignment mapping when reassigning work after a snapshot delete
had gone through. This only affects partial snaphots but essentially
dead-locks the snapshot process.
Closes#61762
System indices can be snapshotted and are therefore potential candidates
to be mounted as searchable snapshot indices. As of today nothing
prevents a snapshot to be mounted under an index name starting with .
and this can lead to conflicting situations because searchable snapshot
indices are read-only and Elasticsearch expects some system indices
to be writable; because searchable snapshot indices will soon use an
internal system index (#60522) to speed up recoveries and we should
prevent the system index to be itself a searchable snapshot index
(leading to some deadlock situation for recovery).
This commit introduces a changes to prevent snapshots to be mounted
as a system index.
This reworks `CardinalityUpperBound` to support precise estimates while
maintaining most of the public API. This will allow us to make more
informed choices about the data structures that we use in aggregations.
None of those interesting choices come as part of this change, but they
are more possible with it.
Backport of #61474.
Part of #46106. Simplify the implementation of deprecation logging by
relying of log4j more completely, and implementing additional behaviour
through custom appenders and filters.
The fact that the data node is already blocked on writing
data files did not guarantee that the cluster state that made
the data node start snapshotting is already applied on master.
This could lead to races where the get snapshots action still
runs based on a state without the snapshot in it, tripping the assertion.
Much safer to handle this by waiting on the non-blocking snapshot create
to return, which guarantees that the CS has been applied on master.
Closes#61541
This commit enhances the verbose output for the
`_ingest/pipeline/_simulate?verbose` api. Specifically
this adds the following:
* the pipeline processor is now included in the output
* the conditional (if) and result is now included in the output iff it was defined
* a status field is always displayed. the possible values of status are
* `success` - if the processor ran with out errors
* `error` - if the processor ran but threw an error that was not ingored
* `error_ignored` - if the processor ran but threw an error that was ingored
* `skipped` - if the process did not run (currently only possible if the if condition evaluates to false)
* `dropped` - if the the `drop` processor ran and dropped the document
* a `processor_type` field for the type of processor (e.g. set, rename, etc.)
* throw a better error if trying to simulate with a pipeline that does not exist
closes#56004
This commit adds the functionality to allocate newly created indices on nodes in the "hot" tier by
default when they are created.
This does not break existing behavior, as nodes with the `data` role are considered to be part of
the hot tier. Users that separate their deployments by using the `data_hot` (and `data_warm`,
`data_cold`, `data_frozen`) roles will have their data allocated on the hot tier nodes now by
default.
This change is a little more complicated than changing the default value for
`index.routing.allocation.include._tier` from null to "data_hot". Instead, this adds the ability to
have a plugin inject a setting into the builder for a newly created index. This has the benefit of
allowing this setting to be visible as part of the settings when retrieving the index, for example:
```
// Create an index
PUT /eggplant
// Get an index
GET /eggplant?flat_settings
```
Returns the default settings now of:
```json
{
"eggplant" : {
"aliases" : { },
"mappings" : { },
"settings" : {
"index.creation_date" : "1597855465598",
"index.number_of_replicas" : "1",
"index.number_of_shards" : "1",
"index.provided_name" : "eggplant",
"index.routing.allocation.include._tier" : "data_hot",
"index.uuid" : "6ySG78s9RWGystRipoBFCA",
"index.version.created" : "8000099"
}
}
}
```
After the initial setting of this setting, it can be treated like any other index level setting.
This new setting is *not* set on a new index if any of the following is true:
- The index is created with an `index.routing.allocation.include.<anything>` setting
- The index is created with an `index.routing.allocation.exclude.<anything>` setting
- The index is created with an `index.routing.allocation.require.<anything>` setting
- The index is created with a null `index.routing.allocation.include._tier` value
- The index was created from an existing source metadata (shrink, clone, split, etc)
Relates to #60848
Runtime fields need to have a SearchLookup available, when building their fielddata implementations, so that they can look up other fields, runtime or not.
To achieve that, we add a Supplier<SearchLookup> argument to the existing MappedFieldType#fielddataBuilder method.
As we introduce the ability to look up other fields while building fielddata for mapped fields, we implicitly add the ability for a field to require other fields. This requires some protection mechanism that detects dependency cycles to prevent stack overflow errors.
With this commit we also introduce detection for cycles, as well as a limit on the depth of the references for a runtime field. Note that we also plan on introducing cycles detection at compile time, so the runtime cycles detection is a last resort to prevent stack overflow errors but we hope that we can reject runtime fields from being registered in the mappings when they create a cycle in their definition.
Note that this commit does not introduce any production implementation of runtime fields, but is rather a pre-requisite to merge the runtime fields feature branch.
This is a breaking change for MapperPlugins that plug in a mapper, as the signature of MappedFieldType#fielddataBuilder changes from taking a single argument (the index name), to also accept a Supplier<SearchLookup>.
Relates to #59332
Co-authored-by: Nik Everett <nik9000@gmail.com>
Errors from bad mappings at index creation are currently logged at DEBUG level, which
can make it difficult to work out what's going on if the index is being auto-created. This
commit ups the log level to INFO for auto-created indices, and includes some more
information in the log message.
Today the `CoordinatorTests` run the publication process as a single
atomic action; however in production it appears possible that another
master may be elected, publish its state, then fail, then we win another
election, all in between the time we sampled our previous cluster state
and started to publish the one we first thought of.
This violates the `assertClusterStateConsistency()` assertion that
verifies the cluster state update event matches the states we actually
published and applied.
This commit adjusts the tests to run the publication process more
asynchronously so as to allow time for this behaviour to occur. This
should eventually result in a reproduction of the failure in #61437 that
will let us analyse what's really going on there and help us fix it.
Today we use `long` to represent the number of parts of a blob. There's
no need for this extra range, it forces us to do some casting elsewhere,
and indeed when snapshotting we iterate over the parts using an `int`
which would be an infinite loop in case of overflow anyway:
for (int i = 0; i < fileInfo.numberOfParts(); i++) {
This commit changes the representation of the number of parts of a blob
to an `int`.
We convert longs to ints using `Math.toIntExact` in places where we're
sure there will be no overflow, but this doesn't explain the intent of
these conversions very well. This commit introduces a dedicated method
for these conversions, and adds an assertion that we never overflow.
This commit removes the tasks module that only existed to define the
tasks result index, `.tasks`, as a system index. The definition for
the tasks results system index descriptor is moved to the
`SystemIndices` class with a check that no other plugin or module
attempts to define an entry with the same source.
Additionally, this change also makes the pattern for the tasks result
index a wildcard pattern since we will need this when the index is
upgraded (reindex to new name and then alias that to .tasks).
Backport of #61540
DeprecationLogger's constructor should not create two loggers. It was
taking parent logger instance, changing its name with a .deprecation
prefix and creating a new logger.
Most of the time parent logger was not needed. It was causing Log4j to
unnecessarily cache the unused parent logger instance.
depends on #61515
backports #58435
Backport to add case insensitive support for regex queries.
Forks a copy of Lucene’s RegexpQuery and RegExp from Lucene master.
This can be removed when 8.7 Lucene is released.
Closes#59235
Splitting DeprecationLogger into two. HeaderWarningLogger - responsible for adding a response warning headers and ThrottlingLogger - responsible for limiting the duplicated log entries for the same key (previously deprecateAndMaybeLog).
Introducing A ThrottlingAndHeaderWarningLogger which is a base for other common logging usages where both response warning header and logging throttling was needed.
relates #55699
relates #52369
backports #55941
* Faster `equals` for `BytesArray` which is nice since with this change we use it for the search cache
* Lighter `StreamInput` for `BytesArray` that should save memory and some indirection relative to the one on the abstract bytes reference
* Lighter `writeTo` implementation
* Build a `BytesArray` instead of a PagedBytesReference whenever possible to save indirection and memory
This is mostly motivated by the performance issues we are seeing around the GET mappings
REST API which (in case of a large number of indices) will create decompressing streams in a hot loop
which takes a significant amount of time for the system calls involved in instantiating deflaters
and inflaters.
Also, this fixes a leaked deflater when deserializing cached repository data.
This method might have materialize all the bytes in a reference into a fresh `byte[]`.
Using the stream is much safer and only trivially more expensive + in most cases we now run the fast path via `BytesArray` anyway.
This optimization is more relevant in the context of CCR. When a node in
the follower cluster leaves, we reallocate the shard-follow tasks on
that node to other nodes. The new tasks will overwhelm the follower
cluster with many put-mapping, update-settings requests, although most
of them are noop. This change detects and optimizes the noop
update-settings requests.
This continues #61301, migrating all of the mappers in `server` to the
new `MapperTestCase` which is nicer than `FieldMapperTestCase` because
it doesn't depend on all of Elasticsearch.
It's unnecessary (and adds one string comparison to every request) to special
case the favicon so I added it as a normal REST handler to simplify the code.
Wrapping a `BytesArray` in a `StreamInput` for deserialization is inefficient.
This forces Jackson to internally buffer (i.e. copy) all bytes from the `BytesArray`
before deserializing, adding overhead for copying the bytes and managing the buffers.
This commit fixes a number of spots where `BytesArray` is the most common type of
`BytesReference` to special case this type and parse it more efficiently.
Also improves parsing `String`s to use the more efficient direct `String` parsing APIs.
Today a remote cluster connection comprises a `PING` and a `REG`
channel. The `PING` channel is only used for health checks between the
elected master and the members of its own cluster, so is unused in a
remote cluster connection. This commit removes this unused connection.
For large responses to the get mappings request, the serialization
to XContent can be extremely slow (serializing mappings is expensive since
we have to decompress and deserialize the mapping source).
To not introduce instability on the IO thread handling the get mappings response
we should move the serialization to the management pool.
The trade-off of introducing one or two new context switches for responses that are
small enough to not cause trouble on the transport thread to prevent instability
in case of a large number of mappings in the cluster seems worth it.
It is not realistic to drop messages without eventually failing.
To retain the coverage of long pauses this PR adjusts the blackholed
behavior to fail a send after 24h (which is assumed to be longer than any
timeout in the system) instead of never.
Closes#61034
Before when a value was copied to a field through a parent field or `copy_to`,
we parsed it using the `FieldMapper` from the source field. Instead we should
parse it using the target `FieldMapper`. This ensures that we apply the
appropriate mapping type and options to the copied value.
To implement the fix cleanly, this PR refactors the value parsing strategy. Now
instead of looking up values directly, field mappers produce a helper object
`ValueFetcher`. The value fetchers are responsible for almost all aspects of
fetching, including looking up the right paths in the _source.
The PR is fairly big but each commit can be reviewed individually.
Fixes#61033.
Previously we didn't retain the requested fields when performing a shallow copy
of the search source. This meant that when a search was rewritten, we could drop
the requested fields and fail to return them in the response.
Saving some cycles here and there on the IO loop:
* Don't instantiate new `Runnable` to execute on `SAME` in a few spots
* Don't instantiate complicated wrapped stream for empty messages
* Stop instantiating almost never used `ClusterStateObserver` in two spots
* Some minor cleanup and preventing pointless `Predicate<>` instantiation in transport master node action
In addition, this commit converts ScaledFloatFieldMapper as it was relying
on a number of static values taken from NumberFieldMapper that had changed
or been removed.
This switches a few tests for field mappers from `ESSingleNodeTestCase`
to `ESTestCase` because, in general, we prefer to avoid
`ESSingleNodeTestCase` when we can because it is slow and "big". "Big"
here means that it pulls in an entire node, making it difficult to
reason about what you are testing.
We have to set the recovery setting to `0` if we don't want throttling
from recoveries. Otherwise the randomized value used for this setting in
tests can lead to throttling unexpectedly.
Closes#61311
With #60683 we stopped forcing aggregating all docs using a single
Aggregator which made some of our accuracy assumptions about the stats
aggregator incorrect. This adds a test that does the forcing and asserts
the old accuracy and adds a test without the forcing with much looser
accuracy guarantees.
Closes#61132
The FieldNamesFieldMapper field has different behaviour for indexes created in
clusters earlier than v6.1, and the code to deal with this was still using the vestigial
FieldType field of FieldMapper in its indexing path. This meant that documents
added after an upgrade were not correctly indexing their field names field. This
commit corrects the parseCreateField method to use the default field type.
Fixes#61305
Today a common reason for a `ShardLockObtainFailedException` is when a
shard is removed from a node and then assigned straight back to it again
before the node has had a chance to shut the previous shard instance
down. For instance, this can happen if a node briefly leaves the cluster
holding a primary with no in-sync replicas.
The message in this case is typically as follows:
obtaining shard lock timed out after 5000ms, previous lock details: [shard creation] trying to lock for [shard creation]
This is pretty hard to interpret, and doesn't raise the important
question: "why didn't the shard shut down sooner?"
With this change we reword the message a bit, report the age of the
shard lock, and adjust the details to report that the lock is held by a
closing shard:
obtaining shard lock for [starting shard] timed out after [5000ms], lock already held for [closing shard] with age [12345ms]
Relates #38807
We have seen a situation where the total search operations are higher
than expected. Unfortunately, we did not have enough info to figure it
out. This commit adds the failures to the error to provide more context
and adjusts the log level in case of failure to debug.
We only work with heap byte buffers at this point and those we can and do unwrap the
`byte[]` ourselves and use `BytesArray` instead of a needless level of indirection via `ByteBuffer`.
There is a corner case here in which during partial snapshot the index is
deleted right between starting the snapshot in the CS and the data node getting to work
on it, causing the data node the fail that shard snapshot and making the snapshot `PARTIAL`.
Closes#61208
Adds a method to make a random date `DateFormatter` pattern. We expect
this'll be useful for runtime fields to compate their formatting with
the standard date field.
Currently we occasionally can get ArithmeticException from parsing bad input
values on 'date' fields that are passed on even if 'ignore_malformed' is set.
This change adds this exception to the ones we already catch for malformed
values.
Closes#52634
Today we allocate a new `byte[]` for each document written to the
cluster state. Some of these documents may be quite large. We need a
buffer that's at least as large as the largest document, but there's no
need to use a fresh buffer for each document.
With this commit we re-use the same `byte[]` much more, only allocating
it afresh if we need a larger one, and using the buffer needed for one
round of persistence as a hint for the size needed for the next one.
The 7.x branch preserves the legacy discovery mechanism from 6.x purely
for running internal cluster tests; this mechanism is otherwise
completely untested and unsupported. However it is still technically
possible to use it outside of the test suite if you dig through the
source code to work out what settings need to be set. With this change
we make it impossible to use this mechanism in production.
Closes#61177
The ReloadSecureSettingsIT makes requests to the reload settings apis.
In 7.x, the client used from the integ test infrastructure may be a
transport client. In that case, the expected exception type, and causes
the test to fail (though it will hang indefinitely due to not counting
down the latch, see
https://github.com/elastic/elasticsearch/pull/60800). This commit adds
unwrapping of the remote exception to get the underlying expected
exception.
closes#51546
Use transport blocking to make relocation take forever instead of relying on the relocation to take long enough to clash with the snapshot.
Closes#61069
It is disastrous if we commit an incremental cluster state update
without having written the full state first. We assert that this doesn't
happen, but it is hard to fully test the myriad ways that things might
fail in a messy production environment. Given the disastrous
consequences it is worth erring on the side of caution in this area.
This commit fails invalid writes even if assertions are disabled.
This commit adds the `data_hot`, `data_warm`, `data_cold`, and `data_frozen` node roles to the
x-pack plugin. These roles are intended to be the base for the formalization of data tiers in
Elasticsearch.
These roles all act as data nodes (meaning shards can be allocated to them). Nodes with the existing
`data` role acts as though they have all of the roles configured (it is a hot, warm, cold, and
frozen node).
This also includes a custom `AllocationDecider` that allows the user to configure the following
settings on a cluster level:
- `cluster.routing.allocation.require._tier`
- `cluster.routing.allocation.include._tier`
- `cluster.routing.allocation.exclude._tier`
And in index settings:
- `index.routing.allocation.require._tier`
- `index.routing.allocation.include._tier`
- `index.routing.allocation.exclude._tier`
Relates to #60848
Converting AllFieldMapper to parametrized form ended up not being run through BWC
testing, resulting in an incorrect implementation being committed. This commit fixes
the serialization, and adds unit tests as well as unmuting the BWC test that uncovered
the bug.
Fixes#60986
Elasticsearch currently blocks writes by default when a master is unavailable. The cluster.no_master_block setting allows
a user to change this behavior to also block reads when a master is unavailable. This PR introduces a way to now also still
allow writes when a master is offline. Writes will continue to work as long as routing table changes are not needed (as
those require the master for consistency), or if dynamic mapping updates are not required (as again, these require the
master for consistency).
Eventually we should switch the default of cluster.no_master_block to this new mode.
Today a snapshot repository verification ensures that all master-eligible and data nodes have write access to the
snapshot repository (and can see each other's data) since taking a snapshot requires data nodes and the currently
elected master to write to the repository. However, a dedicated voting-only master-eligible node is not a data node and
will never be the elected master so we should not require it to have write access to the repository.
Closes#59649
Repositories can't be unregistered when they are actively being used for snapshots or restores. Wildcard repository
deletes could silently bypass the "repo in use" checks however, which is now fixed.
This makes KeywordFieldMapper extend ParametrizedFieldMapper, with explicitly
defined parameters.
In addition, we add a new option to Parameter, restrictedStringParam, which
accepts a restricted set of string options.
The Query string parser was not delegating the construction of wildcard/regex queries to the underlying field type.
The wildcard field has special data structures and queries that operate on them so cannot rely on the basic regex/wildcard queries that were being used for other fields.
Closes#60957
Use thread-local buffers and deflater and inflater instances to speed up
compressing and decompressing from in-memory bytes.
Not manually invoking `end()` on these should be safe since their off-heap memory
will eventually be reclaimed by the finalizer thread which should not be an issue for thread-locals
that are not instantiated at a high frequency.
This significantly reduces the amount of byte copying and object creation relative to the previous approach
which had to create a fresh temporary buffer (that was then resized multiple times during operations), copied
bytes out of that buffer to a freshly allocated `byte[]`, used 4k stream buffers needlessly when working with
bytes that are already in arrays (`writeTo` handles efficient writing to the compression logic now) etc.
Relates #57284 which should be helped by this change to some degree.
Also, I expect this change to speed up mapping/template updates a little as those make heavy use of these
code paths.
This commit introduces a new thread pool, `system_read`, which is
intended for use by system indices for all read operations (get and
search). The `system_read` pool is a fixed thread pool with a maximum
number of threads equal to lesser of half of the available processors
or 5. Given the combination of both get and read operations in this
thread pool, the queue size has been set to 2000. The motivation for
this change is to allow system read operations to be serviced in spite
of the number of user searches.
In order to avoid a significant performance hit due to pattern matching
on all search requests, a new metadata flag is added to mark indices
as system or non-system. Previously created system indices will have
flag added to their metadata upon upgrade to a version with this
capability.
Additionally, this change also introduces a new class, `SystemIndices`,
which encapsulates logic around system indices. Currently, the class
provides a method to check if an index is a system index and a method
to find a matching index descriptor given the name of an index.
Relates #50251
Relates #37867
Backport of #57936
We accept _source values with multiple levels of arrays, such as
`"field": [[[1, 2]]]`. This PR ensures that field retrieval can handle nested
arrays by unwrapping the arrays before parsing the values.
Same as https://github.com/elastic/elasticsearch/pull/43288 for GCS.
We don't need to do the bucket exists check before using the repo, that just needlessly
increases the necessary permissions for using the GCS repository.
Because the 'fields' option loads from _source (which is a stored field), it is
not possible to retrieve 'fields' when stored_fields are disabled.
This also fixes#60912, where setting stored_fields: _none_ prevented the
_ignored fields from being loaded and caused a parsing exception.
This moves the `distance_feature` query building out of
`DistanceFeatureQueryBuilder` and into subclasses of `MappedFieldType`.
Without this we don't have a chance of supporting this for runtime
fields. In general I'm not sad to see the `instanceof`s go.
Co-authored-by: Elastic Machine <elasticmachine@users.noreply.github.com>
This way is faster, saving about 8% on the microbenchmark that rounds to
the nearest month. That is in the hot path for `date_histogram` which is
a very popular aggregation so it seems worth it to at least try and
speed it up a little.
Co-authored-by: Elastic Machine <elasticmachine@users.noreply.github.com>
This commit removes the ability to test the top level result of an aggregator
before it runs the final reduce. All aggregator tests that use AggregatorTestCase#search
are rewritten with AggregatorTestCase#searchAndReduce in order to ensure that we test
the final output (the one sent to the end user) rather than an intermediary result
that could be different.
This change also removes spurious commits triggered on top of a random index writer.
These commits slow down the tests and are redundant with the commits that the
random index writer performs.
There is no point in timing out a join attempt any more once a cluster
is entirely in 7.x. Timing out and retrying with the same master is
pointless, and an in-flight join attempt to one master no longer blocks
attempts to join other masters. This commit deprecates this unnecessary
setting and removes its effect from the joining process.
Relates #60873 which removes this setting in master.
This commit makes IpFieldMapper extend ParametrizedFieldMapper. It also
updates the IpFieldMapper docs to add the ignore_malformed parameter,
which was not previously documented.
Sometimes this test would refresh the disk stats so quickly that it hit
the refresh rate limiter even though it was almost completely disabled.
This commit allows the rate limiter to be completely disabled.
Closes#60587
Several /proc files are expected to contain a single line. We assert on
this in tests, but the contents of the file are lost and the assertion
therefore lacks important information to debug why the file appeared to
have multiple lines. This commit dumps the contents of the file on
assertion failure.
relates #59284
The ReloadSecureSettingsIT uses latches to ensure coordination across
requests to the underlying in memory cluster. However, in the case of an
expected failure, if the assertion fails, the latch will never be
counted down, and will cause the test to hang indefinitely. This commit
ensures the latch is always counted down with a try/finally.
relates #51546
Collapse search queries that sort by a field can throw
an ArrayStoreException due to a bug in the [sort optimization](https://github.com/elastic/elasticsearch/pull/51852)
introduced in 7.7.0. Search collapsing were not supposed to
be eligible for this sort optimization so this change explicitly
filters them from this new feature.
Currently the transport replication action does not propagate the force
execution parameter when acquiring the indexing permit. The logic to
acquire the index permit supports force execution, so this parameter
should be propagate. Fixes#60359.
This pull request adds recovery state tracking for Searchable Snapshots.
In order to track recoveries for searchable snapshot backed indices, this pull
request adds a new type of RecoveryState.
This newRecoveryState instance is able to deal with the
small differences that arise during Searchable snapshots recoveries.
Those differences can be summarized as follows:
- The Directory implementation that's provided by SearchableSnapshots mark the
snapshot files as reused during recovery. In order to keep track of the
recovery process as the cache is pre-warmed, those files shouldn't be marked
as reused.
- Once the shard is created, the cache starts its pre-warming phase, meaning that
we should keep track of those downloads during that process and tie the recovery
to this pre-warming phase. The shard is considered recovered once this pre-warming
phase has finished.
Backport of #60505
This commit uses the new location for the reindex java-api documentation.
Temporary files have been left behind to pacify the docs build.
related #60339
* Stop redundantly creating a `0` length `ByteArray` that is never used
* Add efficient way to get a minimal size copy of the bytes in a `BytesStreamOutput`
* Avoid multiple redundant `byte[]` copies in search cache key creation
Implements license degradation behavior for searchable snapshots. Snapshot-backed shards are failed when the license becomes invalid, and shards won't be reallocated. After valid license is put in place again, shards are allocated again.
Currently, validation of mappers (checking that cross-references are correct, limits on
field name lengths and object depths, multiple definitions, etc) is performed by the
MapperService. This means that any mapper-specific validation, for example that done
on the CompletionFieldMapper, needs to be called specifically from core server code,
and so we can't add validation to mappers that live in plugins.
This commit reworks the validation framework so that mapper-specific validation is
done on the Mapper itself. Mapper gets a new `validate(MappingLookup)`
method (already present on `MetadataFieldMapper` and now pulled up to the parent
interface), which is called from a new `DocumentMapper.validate()` method. All
the validation code currently living on `MapperService` moves either to individual
mapper implementations (FieldAliasMapper, CompletionFieldMapper) or into
`MappingLookup`, an altered `DocumentFieldMappers` which now knows about
object fields and can check for duplicate definitions, or into DocumentMapper
which handles soft limit checks.
In the metadata persistence logic we failed to override the bulk write
method on the FilterOutputStream resulting in all the writes to it
running byte-by-byte in a loop adding a large number of bounds checks
needlessly.
Small oversight in #56078 that only showed up during backporting where a stream copy was turned from a non-closing to a closing one. Enhanced part of a test in this PR to make it show up in master also even though we practically never use this method with stream targets that actually close.
We have various ways of copying between two streams and handling thread-local
buffers throughout the codebase. This commit unifies a number of them and
removes buffer allocations in many spots.
Previously if an inner_hits block required _ source, we would reload and parse
the root document's source for every hit. This PR adds a shared SourceLookup to
the inner hits context that allows inner hits to reuse parsed source if it's
already available. This matches our approach for sharing the root document ID.
Relates to #32818.
This PR simplifies the hierarchy for ordinals field data classes:
* Remove `AbstractIndexFieldData`, since only `AbstractIndexOrdinalsFieldData`
inherits directly from it.
* Make `SortedSetOrdinalsIndexFieldData` extend
`AbstractIndexOrdinalsFieldData`. This lets us remove some redundant code.
Allows nanosecond resolution in search_after (#60328)
This fixes `search_after` to properly parse string formatted dates that
have nanosecond resolution.
Closes#52424
Today when a node fails to properly deserialize a transport message with
a parent task we log the following relatively uninformative message:
java.lang.IllegalStateException: Message not fully read (response) for requestId [9999], handler [org.elasticsearch.transport.TransportService$ContextRestoreResponseHandler/org.elasticsearch.transport.TransportService$ContextRestoreResponseHandler/org.elasticsearch.transport.TransportService$6@abcdefgh], error [false]; resetting
In particular, the wrapping of the listener in the `TransportService`
obscures all clues as to the source of the problem, e.g. the action name
or the identity of the underlying listener. This commit exposes the
inner listener to the logs.
Also if the listener is wrapped with `ContextPreservingActionListener`
then its identity is similarly hidden. This commit also exposes the
wrapped listener in this case.
Relates #38939
Same as #59905 but for shard level metadata. Since we wnat to retain
the ability to do safe+atomic writes for non-uuid shard generations
this PR has to create two separate write paths for both kinds of
shard generations.
Writing the `index.latest` blob is unnecessary unless the contents of the repository
are to be used as a URL-repository. Also, in some edge cases, the fact that `index.latest` is the only
blob in the repository that regularly gets overwritten was causing compatibility issues with
some backing blobstores (Azure no-overwrite policy, Hitachy S3 equivalent).
=> this commit changes behavior to make snapshots not fail if writing `index.latest` fails
and adds a setting to disable writing `index.latest`.
For consistency reasons (and reducing the overload of IllegalArgumentException)
this changes the exception thrown when trying to create a data stream
that already exists.
(cherry picked from commit ac2184c4614bba0f3ee377da49aea0daed98bab4)
Signed-off-by: Andrei Dan <andrei.dan@elastic.co>
This change ensures that we call the listener for partial merge failure **before**
calling the completion listener in order to avoid race condition in tests.
Closes#60446
- Replace immediate task creations by using task avoidance api
- One step closer to #56610
- Still many tasks are created during configuration phase. Tackled in separate steps
The `SourceLookup` class provides access to the _source for a particular
document, specified through `SourceLookup#setSegmentAndDocument`. Previously
the search context contained a single `SourceLookup` that was shared between
different fetch subphases. It was hard to reason about its state: is
`SourceLookup` set to the expected document? Is the _source already loaded and
available?
Instead of using a global source lookup, the fetch hit context now provides
access to a lookup that is set to load from the hit document.
This refactor closes#31000, since the same `SourceLookup` is no longer shared
between the 'fetch _source phase' and script execution.
The oversize algorithm was allocating more pages than necessary to accommodate `minTargetSize`.
An example would be that a 16k page size and 15k `minTargetSize` would result in a new size of 32k (2 pages).
The difference between the minimum number of necessary pages and the estimated size then keeps growing as sizes increase.
I don't think there is much value in preemptively allocating pages by over-sizing aggressively since the behavior of
the system is quite different from that of a single array where over-sizing avoids copying
once the minimum target size is more than a single page.
Relates #60173 which lead me to this when `BytesStreamOutput` would allocate a large number of never used
pages during serialization of repository metadata.
Instances of `BlobStoreIndexShardSnapshots` can be of non-trivial size. In case of snapshotting a larger
number of shards the previous execution order would lead to memory use proportional to the number of shards
for these objects. With this change, the number of these objects on heap is bounded by the size of the snapshot
pool (except for in the BwC format path).
This PR makes it so that they are written to the repository at the earliest possible point in time
so that they can be garbage collected.
If shard generations are used, we can safely write these right at the beginning of the shard snapshot.
If shard generations are not used we can only write them at the end of the shard snapshot after all
other blobs have been written.
Closes#60173
CCR will stop functioning if the master node is on 7.8, but data nodes
are before that version because the master node considers that all data
nodes do not have the remote cluster client role. This commit allows CCR
work on data nodes with legacy roles only.
Relates #54146
Relates #59375
This fixes the behavior of the snapshot state machine in the following edge case:
1. Snapshot is running
2. Delete/abort for the snapshot is started
3. Snapshot fails to finalize
We were not removing the failed snapshot id from the list of snapshots to delete in the delete.
This lead to an error in the repository, which throws if we try to delete a non-existing snapshot.
This commmit updates the deletions in progress by removing the failed snapshot id.
The fact that this could lead to snapshot delete entries without any snapshot ids is not optimized
on purpose because it allows for another attempt at writing clean `RepositoryData` and will run basic
cleanup on the repository (root level blobs and stale indices) and thus bring the repository back into
a clean state after a failed finalization.
Closes#60274
The test assumed that the master fail-over would always work out as a single step.
This is not guaranteed however and we can randomly see master failing over twice,
in which case the transport listener will be failed on the node that stops being
leader and we have to catch an exception for the deletes as well just like we do
for the snapshot.
Closes#60262
Adds a full list of supported aggregations to the node info API. This list
will be used in transform tests and telemetry mapping tests that will be added
as follow-up PRs.
Fixes#59774
This feature adds a new `fields` parameter to the search request, which
consults both the document `_source` and the mappings to fetch fields in a
consistent way. The PR merges the `field-retrieval` feature branch.
Addresses #49028 and #55363.
Transport connections between nodes remain in place until one or other
node shuts down or the connection is disrupted by a flaky network.
Today it is very difficult to demonstrate that transient failures and
cluster instability are caused by the network even though this is often
the case. In particular, transport connections open and close without
logging anything, even at `DEBUG` level, making it very hard to quantify
the scale of the problem or to correlate the networking problems with
external events.
This commit adds the missing `DEBUG`-level logging when transport
connections open and close, and also tracks the total number of
transport connections a node has opened as a measure of the stability of
the underlying network.
* Fix Test Failure in testCorrectCountsForDoneShards
Fixing the freak edge case where the node shard status request returns before
the node was able to send the state update request to master and update the cluster state.
Without this change, the snapshot shard status would report as `DONE` once the data node
has finished updating the shard in the cluster state.
If the data node then drops out of the cluster before the state has been updated, then
the status will jump to "FAILURE" because the master updates the state once the data node
leaves the cluster.
Closes#60247
Today if a cluster state observer's listener takes a long time to
process a notification then we log the following rather useless warning
message:
[notifying listener [org.elasticsearch.cluster.ClusterStateObserver$ObserverClusterStateListener@12345678]] took [34567ms]
This commit adds a handful of simple `toString()` implementations in
order to identify the owner of the listener in question.
This change forks the execution of partial
reduces in the coordinating node to the search thread pool.
It also ensures that partial reduces are executed sequentially
and asynchronously in order to limit the memory and cpu that a
single search request can use but also to avoid blocking a
network thread.
If a partial reduce fails with an exception, the search
request is cancelled and the reporting of the error is
delayed to the start of the fetch phase (when the final
reduce is performed). This ensures that we cleanup the
in-flight search requests before returning an error to
the user.
Closes#53411
Relates #51857
We can save one round of serializing `RepositoryData` on the write path.
This also leads to somewhat better compression because we compress larger chunks
in one go potentially when compared to serializing and compressing in one go.
Also, fixed the double wrapping of collections when copying the repository
data instance via the `withGenId`.
Keepalive options are not well-documented (only in transport section, although also available at http and network level).
Co-authored-by: David Turner <david.turner@elastic.co>
Co-authored-by: James Rodewig <40268737+jrodewig@users.noreply.github.com>
keepalives tell any intermediate devices that the connection remains alive, which helps with overzealous firewalls that are
killing idle connections. keepalives are enabled by default in Elasticsearch, but use system defaults for their
configuration, which often times do not have reasonable defaults (e.g. 7200s for TCP_KEEP_IDLE) in the context of
distributed systems such as Elasticsearch.
This PR sets the socket-level keep_alive options for network.tcp.{keep_idle,keep_interval} to 5 minutes on configurations
that support it (>= Java 11 & (MacOS || Linux)) and where the system defaults are set to something higher than 5
minutes. This helps keep the connections alive while not interfering with system defaults or user-specified settings
unless they are deemed to be set too high by providing better out-of-the-box defaults.
Currently, we do not categorize the operation type in the rejection
exception messsage when we reject an indexing operation for indexing
memory limits. This commit fixes this to ensure that it is identified as
coordinating, primary, or replica.
Currently the logic to rejection replica rejections is evaluate before
adding the additional bytes of the current operation. This means that
the first replica operation which should be rejected will be allowed to
proceed. This commit fixes this logic and adds unit level test to ensure
indexing pressure behavior is correct.
When a document which is distant from existing buckets gets collected, the
`variable_width_histogram` will create a new bucket and then insert it
into the ordered list of buckets.
Currently, a new merge map array is created to move this bucket. This is
very expensive as there might be thousands of buckets.
This PR creates `mergeBuckets(UnaryOperator<Long> mergeMap)` methods in
`BucketsAggregator` and `MergingBucketsDefferingCollector`, and updates
the `variable_width_histogram` to use them. This eliminates the need to
create an entire merge map array for each new bucket and reduces the
memory overhead of the algorithm.
Co-authored-by: James Dorfman <jamesdorfman@users.noreply.github.com>
In almost all cases we write uuid named files via this method.
Preemptively deleting just wastes IO ops, we can delete after a write failed
and retry the write to cover the few cases where we actually do an overwrite.
Putting an ingest pipeline used to require that the user calling
it had permission to get nodes info as well as permission to
manage ingest. This was due to an internal implementaton detail
that was not visible to the end user.
This change alters the behaviour so that a user with the
manage_pipeline cluster privilege can put an ingest pipeline
regardless of whether they have the separate privilege to get
nodes info. The internal implementation detail now runs as
the internal _xpack user when security is enabled.
Backport of #60106
There is a very unlikely but possible test failure in this test.
The `SnapshotsService` continues iterating over queued operations after
resolving the transport listener. This can lead to a situation where the moved
repository data is not picked up when running the delete (even though we
have the concurrent modifications BwC mode activated) concurrently.
I fixed this in the test so that the test still verifies that this setting works.
Technically speaking, one could add logic to the way we queue and execute repo operations
to address this special case. Since this case only comes about with the concurrent modifications
setting enabled (and the setting is gone in master already) I don't really see a reason to improve
the logic here since we should always fail queued up repo operations on concurrent modification for
safety reasons.
Today there is a node-level `canAllocate` override which the balancer
uses to ignore certain nodes to which it is certain no more shards can
be allocated. In fact this override only ignores nodes which have hit
the rarely-used `cluster.routing.allocation.total_shards_per_node`
limit, so this optimization doesn't have a meaningful impact on real
clusters.
This commit removes this unnecessary fast path from the balancer, and
also removes all the machinery needed to support it.
The single element `close` calls go through a very inefficient path that includes creating
a one element list.
`releaseOnce` is only with a single non-null input in production in two spots so no need for
varargs and any complexity here.
`ReleasableBytesStreamOutput` does not require any `releaseOnce` wrapping because we already have
that kind of logic implemented in `org.elasticsearch.common.util.AbstractArray` (which we were
wrapping here) already.
This improves modularity and also fixes some issues when `docvalues_fields` is
used within `inner_hits` or the `top_hits` agg:
* We previously didn't resolve wildcards in field names.
* We also forgot to enforce the limit `index.max_docvalue_fields_search`.
Due to complicated access checks (reads and writes execute in their own access context) on some repositories (GCS, Azure, HDFS), using a hard coded buffer size of 4k for restores was needlessly inefficient.
By the same token, the use of stream copying with the default 8k buffer size for blob writes was inefficient as well.
We also had dedicated, undocumented buffer size settings for HDFS and FS repositories. For these two we would use a 100k buffer by default. We did not have such a setting for e.g. GCS though, which would only use an 8k read buffer which is needlessly small for reading from a raw `URLConnection`.
This commit adds an undocumented setting that sets the default buffer size to `128k` for all repositories. It removes wasteful allocation of such a large buffer for small writes and reads in case of HDFS and FS repositories (i.e. still using the smaller buffer to write metadata) but uses a large buffer for doing restores and uploading segment blobs.
This should speed up Azure and GCS restores and snapshots in a non-trivial way as well as save some memory when reading small blobs on FS and HFDS repositories.
This commit continues on the work in #59801 and makes other
implementors of the LocalNodeMasterListener interface thread safe in
that they will no longer allow the callbacks to run on different
threads and possibly race each other. This also helps address other
issues where these events could be queued to wait for execution while
the service keeps moving forward thinking it is the master even when
that is not the case.
In order to accomplish this, the LocalNodeMasterListener no longer has
the executorName() method to prevent future uses that could encounter
this surprising behavior.
Each use was inspected and if the class was also a
ClusterStateListener, the implementation of LocalNodeMasterListener
was removed in favor of a single listener that combined the logic. A
single listener is used and there is currently no guarantee on execution
order between ClusterStateListeners and LocalNodeMasterListeners,
so a future change there could cause undesired consequences. For other
classes, the implementations of the callbacks were inspected and if the
operations were lightweight, the overriden executorName method was
removed to use the default, which runs on the same thread.
Backport of #59932
Sometimes there is the need to make a field required in the mappings, and validate that a value has been provided for it. This can be done through a validator when using ParametrizedFieldMapper, but validators need to run also when a value for a field has not been specified.
Relates to #59332
There were two subtle bugs here from backporting #56911 to 7.x.
1. We passed `null` for the `shards` map which isn't nullable any longer
when creating `SnapshotsInProgress.Entry`, fixed by just passing an empty map
like the `null` handling did in the past.
2. The removal of a failed `INIT` state snapshot from the cluster state tried
removing it from the finalization loop (the set of repository names that are
currently finalizing). This will trip an assertion since the snapshot failed
before its repository was put into the set. I made the logic ignore the set
in case we remove a failed `INIT` state snapshot to restore the old logic to
exactly as it was before the concurrent snapshots backport to be on the safe
side here.
Also, added tests that explicitly call the old code paths because as can be seen
from initially missing this, the BwC tests will only run in the configuration new
version master, old version nodes ever so often and having a deterministic test
for the old state machine seems the safest bet here.
Closes#59986
For ingest node processors a per processor description
was recently added. This commit displays that description
in the verbose output of the pipeline simulation.
related #57906
In #54716 I removed pipeline aggregators from the aggregation result
tree and caused us to read them from the request. This saves a bunch of
round trip bytes, which is neat. But there was a bug in the backwards
compatibility logic. You see, we still have to give the pipeline
aggregations to nodes older than 7.8 over the wire because that is how
they know what pipelines to run. They have the pipelines in the request
but they don't read them. They use the ones in the response tree.
Anyway, we had a bug where we were never sending pipelines defined two
levels down. So while you are upgrading the pipeline wouldn't run.
Sometimes. If the data node of the "first" result was post-7.8 and the
coordinating node was pre-7.8.
This fixes the bug.
Today the `InternalClusterInfoService` uses the
`LocalNodeMasterListener` interface to start/stop its operations. Since
the `onMaster` and `offMaster` methods are called on the `MANAGEMENT`
threadpool, there's no guarantee that they run in the correct sequence,
which could result in an elected master failing to regularly update the
cluster info.
Since this service is also a `ClusterStateListener` we may as well drop
the usage of the `LocalNodeMasterListener` interface and simply update
the status of the local node on the applier thread in `clusterChanged`
to ensure consistency.
Additionally, today the `InternalClusterInfoService` uses a simple flag
to track whether the local node is the elected master or not. If the
node stops being the master and then starts again within a few seconds
then the scheduled updates from the old mastership might carry on
running in addition to the ones for the new mastership.
This commit addresses that by tracking the identity of the scheduled
update job and creating a new job for each mastership.
The TypeParser implementations of all ParametrizedFieldMapper descendant classes are
essentially the same - stateless, requiring the construction of a Builder object, and calling
parse on it before returning it. We can make this easier (and less error-prone) to
implement by wrapping the logic up into a final class, which takes a function to produce
the Builder from a name and parser context.
We never used the `IndexSettings` parameter and we only used the
`MappedFieldType` parameter to get the name of the field which we
already know everywhere where we build the `IFD.Builder`. This allows us
to drop a fair bit of ceremony from a couple of tests.
ParametrizedFieldMapper overrides `toXContent` from `FieldMapper`, yet it could override `doXContentBody` and rely on the `toXContent` from the base class. Additionally, this allows to make `doXContentBody` final. Also, toXContent is still overridden only to make it final.
With uuid named segment data blobs there is no reason to ensure no overwrites are happening
for these blobs when writing. On the contrary, at least on Azure this check can conflict with
the SDK's retrying and cause upload failures randomly.
Refactored `CheckSumBlobStoreFormat` so it can more easily be reused in
other functionality (i.e. upcoming repair logic).
Simplified away constant `failIfAlreadyExists` parameter and removed the atomic
write method and its tests.
The atomic write method was only used in a single spot and that spot has now been adjusted to
work the same way writing root level metadata works.
Follow up to #59606 using some of the new infrastructure and making similar cleanups (and due to at times better handling of size hints and empty collections also optimizations in the stream utility methods this also means speedups) in various spots in the core codebase.
Previously we constructed a GeometryFormat object and delegated point parsing to
it. This wasn't a good fit conceptually because each GeometryFormat instance
didn't represent a distinct point format.
This replaces that data structure that we use to resolve bucket ids in
bucketing aggs that are inside other bucketing aggs. This replaces the
"legoed together" data structure with a purpose built `LongLongHash`
with semantics similar to `LongHash`, except that it has two `long`s
as keys instead of one.
The microbenchmarks show a fairly substantial performance gain on the
hot path, around 30%. Rally's higher level benchmarks show anywhere
from 0 to 7% speed improvements. Not as much as I'd hoped, but nothing
to sneeze at. And, after all, we all allocating slightly less data per
owningBucketOrd, which is always nice.
Today `GET _nodes/stats/fs` includes `{least,most}_usage_estimate`
fields for some nodes. These fields have rather strange semantics. They
are only reported on the elected master and on nodes that have been the
elected master since they were last restarted; when a node stops being
the elected master these stats remain in place but we stop updating them
so they may become arbitrarily stale.
This means that these statistics are pretty meaningless and impossible
to use correctly. Even if they were kept up to date they're never
reported for data-only nodes anyway, despite the fact that data nodes
are the ones where we care most about disk usage. The information needed
to compute the path with the least/most available space is already
provided in the rest the stats output, so we can treat the inclusion of
these stats as a bug and fix it by simply removing them in this commit.
Since these stats were always optional and mostly omitted (for opaque
reasons) this is not considered a breaking change.
This cleans up a few rough edged in the `variable_width_histogram`,
mostly found by @wwang500:
1. Setting its tuning parameters in an unexpected order could cause the
request to fail.
2. We checked that the maximum number of buckets was both less than
50000 and MAX_BUCKETS. This drops the 50000.
3. Fixes a divide by 0 that can occur of the `shard_size` is 1.
4. Fixes a divide by 0 that can occur if the `shard_size * 3` overflows
a signed int.
5. Requires `shard_size * 3 / 4` to be at least `buckets`. If it is less
than `buckets` we will very consistently return fewer buckets than
requested. For the most part we expect folks to leave it at the
default. If they change it, we expect it to be much bigger than
`buckets`.
6. Allocate a smaller `mergeMap` in when initially bucketing requests
that don't use the entire `shard_size * 3 / 4`. Its just a waste.
7. Default `shard_size` to `10 * buckets` rather than `100`. It *looks*
like that was our intention the whole time. And it feels like it'd
keep the algorithm humming along more smoothly.
8. Default the `initial_buffer` to `min(10 * shard_size, 50000)` like
we've documented it rather than `5000`. Like the point above, this
feels like the right thing to do to keep the algorithm happy.
Co-authored-by: Elastic Machine <elasticmachine@users.noreply.github.com>
Co-authored-by: Elastic Machine <elasticmachine@users.noreply.github.com>
There was a bug in the geoshape circuit-breaker check where the
hash values array was being allocated before its new size was
accounted for by the circuit breaker.
Fixes#57847.
* Adding new `require_alias` option to indexing requests (#58917)
This commit adds the `require_alias` flag to requests that create new documents.
This flag, when `true` prevents the request from automatically creating an index. Instead, the destination of the request MUST be an alias.
When the flag is not set, or `false`, the behavior defaults to the `action.auto_create_index` settings.
This is useful when an alias is required instead of a concrete index.
closes https://github.com/elastic/elasticsearch/issues/55267
Currently, when we apply a cluster state change to a shard on a non-master node,
we check to see if the mappings need to be updated by comparing the decompressed
serialized mappings from the update against the serialized version of the shard's
existing mappings. However, we already have a much simpler way of checking this,
by comparing mapping versions on the index metadata of the old and new states.
This commit adds a shortcut to MapperService.updateMappings() that compares
these mapping versions, and ignores the merge if they are equal.
This commit makes DateFieldMapper extend ParametrizedFieldMapper,
declaring its parameters explicitly. As well as changes to DateFieldMapper
itself, there are some changes to dynamic mapping code to ensure that
dynamically detected date formats are passed through to new date mapper
builders.
* Fix handling of final pipelines when destination is changed (#59522)
This change fixes final pipelines if destination index is changed during pipeline run:
-final pipelines can't change destination anymore, exception is thrown if they try to
-if request/default pipeline changes destination final pipeline from old index won't be executed
-if request/default pipeline changes destination and new index has final pipeline it will be executed
-default pipeline from new index won't be executed
Additionally TransportBulkAction.resolvePipelines was moved to IngestService as it's needed for resolving pipelines from new index. Tests were moved accordingly.
Closes#57968
After the work to retry transient replication failures, the local and
global checkpoint test metadata can be incremented on a different thread
than the test thread. This appears to introduce an extremely rare
scenario where this data is not visible for later test assertions. This
commit fixes the issue by using synchronized maps.
Adds a hard_bounds parameter to explicitly limit the buckets that a histogram
can generate. This is especially useful in case of open ended ranges that can
produce a very large number of buckets.
This adds a number of new optional parameters to Parameter, including:
* custom serialization (to handle analyzers)
* deprecated parameter names
* parameter validation
* allowing default values to be based on the values of other parameters
We preserve the previous serialization format of CompletionFieldMapper,
always emitting most fields, in order to meet mapping checks in mixed
version clusters, where the mapper service will check that mappings have
been correctly parsed and updated by checking their serialized outputs.
Eclipse was confused by #59583. It can't see a the public inner
interface within the superclass. This time. Usually that is fine, but
the Eclipse gods don't like this particular code, I guess.
Alan Woodward
Igor Motov
Tim Brooks
Armin Braun
Tanguy Leroux
Dan Hermann
Boice Huang
Nik Everett
Adam Locke
Jason Tedor
Rory Hunter
Jake Landis
Lee Hinman
Luca Cavanna
David Turner
Jay Modi
Przemyslaw Gomulka
markharwood
Nhat Nguyen
Julie Tibshirani
Rory Hunter
Dimitrios Liappis
Mark Tozzi
Christoph Büscher
Ryan Ernst
Yannick Welsch
Jim Ferenczi
Martijn van Groningen
Francisco Fernández Castaño
Andrei Dan
Rene Groeschke
Zachary Tong
James Rodewig
Howard
weizijun
Adrien Grand
David Roberts
Ignacio Vera
Stéphane Campinas
Tal Levy
Benjamin Trent
Przemko Robakowski