Adding an example of how to re-implement the polish stempel analyzer
in case a user want to modify or extend it. In order for the analyzer to be
able to use polish stopwords, also registering a polish_stop filter for the
stempel plugin.
Closes#13150
We had this as a dependency for legacy dependencies that still needed
the Log4j 1.2 API. This appears to no longer be necessary, so this
commit removes this artifact as a dependency.
To remove this dependency, we had to fix a few places where we were
accidentally relying on Log4j 1.2 instead of Log4j 2 (easy to do, since
both APIs were on the compile-time classpath).
Finally, we can remove our custom Netty logger factory. This was needed
when we were on Log4j 1.2 and handled logging in our own unique
way. When we migrated to Log4j 2 we could have dropped this
dependency. However, even then Netty would still pick up Log4j 1.2 since
it was on the classpath, thus the advantage to removing this as a
dependency now.
* Cleanup Bulk Delete Exception Logging
* Follow up to #41368
* Collect all failed blob deletes and add them to the exception message
* Remove logging of blob name list from caller exception logging
* Add Infrastructure to Run 3rd Party Repository Tests
* Add infrastructure to run third party repository tests using our standard JUnit infrastructure
* This is a prerequisite of #42189
* Remove Delete Method from BlobStore (#41619)
* The delete method on the blob store was used almost nowhere and just duplicates the delete method on the blob containers
* The fact that it provided for some recursive delete logic (that did not behave the same way on all implementations) was not used and not properly tested either
Both of these classes are basically a bloated wrapper around a simple
construct that can simply be a DirectoryFactory interface. This change
removes both classes and replaces them with a simple stateless interface
that creates a new `Directory` per shard. The concept of `index.store` is preserved
since it makes sense from a configuration perspective.
This commit updates the default ciphers and TLS protocols that are used
when the runtime JDK supports them. New cipher support has been
introduced in JDK 11 and 12 along with performance fixes for AES GCM.
The ciphers are ordered with PFS ciphers being most preferred, then
AEAD ciphers, and finally those with mainstream hardware support. When
available stronger encryption is preferred for a given cipher.
This is a backport of #41385 and #41808. There are known JDK bugs with
TLSv1.3 that have been fixed in various versions. These are:
1. The JDK's bundled HttpsServer will endless loop under JDK11 and JDK
12.0 (Fixed in 12.0.1) based on the way the Apache HttpClient performs
a close (half close).
2. In all versions of JDK 11 and 12, the HttpsServer will endless loop
when certificates are not trusted or another handshake error occurs. An
email has been sent to the openjdk security-dev list and #38646 is open
to track this.
3. In JDK 11.0.2 and prior there is a race condition with session
resumption that leads to handshake errors when multiple concurrent
handshakes are going on between the same client and server. This bug
does not appear when client authentication is in use. This is
JDK-8213202, which was fixed in 11.0.3 and 12.0.
4. In JDK 11.0.2 and prior there is a bug where resumed TLS sessions do
not retain peer certificate information. This is JDK-8212885.
The way these issues are addressed is that the current java version is
checked and used to determine the supported protocols for tests that
provoke these issues.
We have a number of places in analysis-handling code where we check
if a field type is a keyword field, and if so then extract the normalizer rather
than pulling the index-time analyzer. However, a keyword normalizer is
really just a special case of an analyzer, so we should be able to simplify this
by setting the normalizer as the index-time analyzer at construction time.
testclusters detect from settings that security is enabled
if a user is not specified using the DSL introduced in this PR, a default one is created
the appropriate wait conditions are used authenticating with the first user defined in the DSL ( or the default user ).
an example DSL to create a user is user username:"test_user" password:"x-pack-test-password" role: "superuser" all keys are optional and default to the values shown in this example
Upgrades the AWS SDK to the same version that we're using for the repository-s3 plugin, providing
testing capabilities to override certain SDK endpoints in order to point them to localhost for testing.
Adds tests for the various credential providers.
Today Elasticsearch accepts, but silently ignores, port ranges in the
`discovery.seed_hosts` setting:
```
discovery.seed_hosts: 10.1.2.3:9300-9400
```
Silently ignoring part of a setting like this is trappy. With this change we
reject seed host addresses of this form.
Closes#40786
Backport of #41404
This is related to #27260. Currently we have a single read buffer that
is no larger than a single TLS packet. This prevents us from reading
multiple TLS packets in a single socket read call. This commit modifies
our TLS work to support reading similar to the plaintext case. The data
will be copied to a (potentially) recycled TLS packet-sized buffer for
interaction with the SSLEngine.
This fixes#41794. Currently the read timeout test queues up responses
in the netty pipeline. These responses are immediately returned in the
write call, but they are not released. This commit releases the
responses. This will cause the leak detector to quit throwing
exceptions.
This is related to #27260. Currently there is a setting
http.read_timeout that allows users to define a read timeout for the
http transport. This commit implements support for this functionality
with the transport-nio plugin. The behavior here is that a repeating
task will be scheduled for the interval defined. If there have been
no requests received since the last run and there are no inflight
requests, the channel will be closed.
* Implement Bulk Deletes for GCS Repository (#41368)
* Just like #40322 for AWS
* We already had a bulk delete API but weren't using it from the blob container implementation, now we are using it
* Made the bulk delete API also compliant with our interface that only suppresses errors about non existent blobs by stating failed deletes (I didn't use any bulk stat action here since having to stat here should be the exception anyway and it would make error handling a lot more complex)
* Fixed bulk delete API to limit its batch size to 100 in line with GCS recommendations
back port of #41368
This is related to #27260. Currently for the SSLDriver we allocate a
dedicated network write buffer and encrypt the data into that buffer one
buffer at a time. This requires constantly switching between encrypting
and flushing. This commit adds a dedicated outbound buffer for SSL
operations that will internally allocate new packet sized buffers as
they are need (for writing encrypted data). This allows us to totally
encrypt an operation before writing it to the network. Eventually it can
be hooked up to buffer recycling.
This commit also backports the following commit:
Handle WRAP ops during SSL read
It is possible that a WRAP operation can occur while decrypting
handshake data in TLS 1.3. The SSLDriver does not currently handle this
well as it does not have access to the outbound buffer during read call.
This commit moves the buffer into the Driver to fix this issue. Data
wrapped during a read call will be queued for writing after the read
call is complete.
Motivated by slow snapshot deletes reported in e.g. #39656 and the fact that these likely are a contributing factor to repositories accumulating stale files over time when deletes fail to finish in time and are interrupted before they can complete.
* Makes snapshot deletion async and parallelizes some steps of the delete process that can be safely run concurrently via the snapshot thread poll
* I did not take the biggest potential speedup step here and parallelize the shard file deletion because that's probably better handled by moving to bulk deletes where possible (and can still be parallelized via the snapshot pool where it isn't). Also, I wanted to keep the size of the PR manageable.
* See https://github.com/elastic/elasticsearch/pull/39656#issuecomment-470492106
* Also, as a side effect this gives the `SnapshotResiliencyTests` a little more coverage for master failover scenarios (since parallel access to a blob store repository during deletes is now possible since a delete isn't a single task anymore).
* By adding a `ThreadPool` reference to the repository this also lays the groundwork to parallelizing shard snapshot uploads to improve the situation reported in #39657
This is related to #27260. Currently for the SSLDriver we allocate a
dedicated network write buffer and encrypt the data into that buffer one
buffer at a time. This requires constantly switching between encrypting
and flushing. This commit adds a dedicated outbound buffer for SSL
operations that will internally allocate new packet sized buffers as
they are need (for writing encrypted data). This allows us to totally
encrypt an operation before writing it to the network. Eventually it can
be hooked up to buffer recycling.
* The check doesn't add much if anything practically, since the S3 repository is eventually consistent and we only log the non-existence of a blob anyway
* We don't do the check on writes for this very reason and documented it as such
* Removing the check saves one API call per single delete speeding up the deletion process and lowering costs
* Adds Bulk delete API to blob container
* Implement bulk delete API for S3
* Adjust S3Fixture to accept both path styles for bulk deletes since the S3 SDK uses both during our ITs
* Closes#40250
* Replace usages RandomizedTestingTask with built-in Gradle Test (#40978)
This commit replaces the existing RandomizedTestingTask and supporting code with Gradle's built-in JUnit support via the Test task type. Additionally, the previous workaround to disable all tasks named "test" and create new unit testing tasks named "unitTest" has been removed such that the "test" task now runs unit tests as per the normal Gradle Java plugin conventions.
(cherry picked from commit 323f312bbc829a63056a79ebe45adced5099f6e6)
* Fix forking JVM runner
* Don't bump shadow plugin version
This change updates our version of httpclient to version 4.5.8, which
contains the fix for HTTPCLIENT-1968, which is a bug where the client
started re-writing paths that contained encoded reserved characters
with their unreserved form.
The xlint exclusions of the following plugins were removed:
* ingest-attachment.
* mapper-size.
* transport-nio. Removing the -try exclusion required some work, because
the NettyAdaptor implements AutoCloseable and NettyAdaptor#close() method
could throw an InterruptedException (ChannelFuture#await() and a generic
Exception is re-thrown, which maybe an ChannelFuture). The easiest way
around this to me seemed that NettyAdaptor should not implement AutoCloseable,
because it is not directly used in a try-with-resources statement.
Relates to #40366
This change adds the following internal refactorings:
* wraps input analyzers into an unmodifiable map in IndexAnalyzers ctor
* removes duplicated indexSetting in IndexAnalyzers
* removes references to IndexAnalyzers from DocumentMapperParser and TypeParser.ParserContext.
It can always be retrieve it from MapperService directly in those cases
The hdfs-fixture is actually executed in plugin/repository-hdfs as a
dependency. The fixture is not needed and actually causes a failure
because we have two copies now and both use the same ports.
* Fix 3rd pary S3 tests
This is allready excluded on line 186, by doing this again here, the
other exclusion from arround that line are removed causing the tests to
fail.
* Fix blacklisting with the fixture
Replaces the vagrant based kerberos fixtures with docker based test fixtures plugin.
The configuration is now entirely static on the docker side and no longer driven by Gradle,
also two different services are being configured since there are two different consumers of the fixture that can run in parallel and require different configurations.
Adds the search_as_you_type field type that acts like a text field optimized
for as-you-type search completion. It creates a couple subfields that analyze
the indexed terms as shingles, against which full terms are queried, and a
prefix subfield that analyze terms as the largest shingle size used and
edge-ngrams, against which partial terms are queried
Adds a match_bool_prefix query type that creates a boolean clause of a term
query for each term except the last, for which a boolean clause with a prefix
query is created.
The match_bool_prefix query is the recommended way of querying a search as you
type field, which will boil down to term queries for each shingle of the input
text on the appropriate shingle field, and the final (possibly partial) term
as a term query on the prefix field. This field type also supports phrase and
phrase prefix queries however
This commit adds an InboundHandler to handle inbound message processing.
With this commit, this code is moved out of the TcpTransport.
Additionally, finer grained unit tests are added to ensure that the
inbound processing works as expected
Currently there are some components of message serializer and sending
that still occur in TcpTransport. This commit makes it possible to
send a message without the TcpTransport by moving all of the remaining
application logic to the OutboundHandler. Additionally, it adds unit
tests to ensure that this logic works as expected.
* Add support for setting and keystore settings
* system properties and env var config
* use testclusters for repository-s3
* Some cleanup of the build.gradle file for plugin-s3
* add runner {} to rest integ test task
FilterDirectory.getPendingDeletions does not delegate, fixed
temporarily by overriding in StoreDirectory.
This in turn caused duplicate file name use after a trimUnsafeCommits
had been done, since a new IndexWriter would not consider the pending
deletes in IndexFileDeleter. This should only happen on windows (AFAIK).
Reenabled doing index updates for all tests using
IndexShardTests.indexOnReplicaWithGaps (which could fail due to above
when using mocked WindowsFS).
Added getPendingDeletions delegation to all elasticsearch
FilterDirectory subclasses that were not trivial test-only overrides to
minimize the risk of hitting this issue in another case.
Currently, we maintain a transport name ("mock-nio", "nio", "netty")
that is passed to a `TcpTransportChannel` when a request is received.
The value of this name is to associate with the task when we register a
task with the task manager. However, it is only possible to run ES with
one transport, so having an implementation specific name is unnecessary.
This commit removes the name and replaces it with the generic
"transport".
This commit adds a variant for every official distribution that omits
the bundled jdk. The "no-jdk" naming is conveyed through the package
classifier, alongside the platform. Package tests are also added for
each new distribution.
This change marks the transport-nio plugin as having a client jar. The
nio transport can be used from a transport client and the
x-pack-transport artifact depends on the transport-nio jar but this jar
is not published. This change marks the transport-nio project as having
a client jar so that the jar may be published in the same way that we
publish the netty4 transport artifact.
The change to actually publish the jar will be handled separately as an
update to the release manager.
* Optimize Bulk Message Parsing and Message Length Parsing
* findNextMarker took almost 1ms per invocation during the PMC rally track
* Fixed to be about an order of magnitude faster by using Netty's bulk `ByteBuf` search
* It is unnecessary to instantiate an object (the input stream wrapper) and throw it away, just to read the `int` length from the message bytes
* Fixed by adding bulk `int` read to BytesReference
* Un-mute and fix BuildExamplePluginsIT
There doesn't seem to be anything wrong with the test iteself.
I think the failure were CI performance related, but while it was muted,
some failures managed to sneak in.
Closes#38784
* PR review
Backport of #39350
Contains the following:
* LUCENE-8635: Move terms dictionary off-heap for non-primary-key fields in `MMapDirectory`
* LUCENE-8292: `TermsEnum` is fully abstract
* LUCENE-8679: Return WITHIN in `EdgeTree#relateTriangle` only when polygon and triangle share one edge
* LUCENE-8676: Nori tokenizer deals correctly with large buffers
* LUCENE-8697: `GraphTokenStreamFiniteStrings` better handles side paths with gaps
* LUCENE-8664: Add `equals` and `hashCode` to `TotalHits`
* LUCENE-8660: `TopDocsCollector` returns accurate hit counts if the total equals the threshold
* LUCENE-8654: `Polygon2D#relateTriangle` fix for when the polygon is inside the triangle
* LUCENE-8645: `Intervals#fixField` can merge intervals from different fields
* LUCENE-8585: Create jump-tables for DocValues at index time
Blob store compression was not enabled for some of the files in
snapshots due to constructor accessing sub-class fields. Fixed to
instead accept compress field as constructor param. Also fixed chunk
size validation to work.
Deprecated repositories.fs.compress setting as well to be able to unify
in a future commit.
DfsPhase captures terms used for scoring a query in order to build global term statistics across
multiple shards for more accurate scoring. It currently does this by building the query's `Weight`
and calling `extractTerms` on it to collect terms, and then calling `IndexSearcher.termStatistics()`
for each collected term. This duplicates work, however, as the various `Weight` implementations
will already have collected these statistics at construction time.
This commit replaces this round-about way of collecting stats, instead using a delegating
IndexSearcher that collects the term contexts and statistics when `IndexSearcher.termStatistics()`
is called from the Weight.
It also fixes a bug when using rescorers, where a `QueryRescorer` would calculate distributed term
statistics, but ignore field statistics. `Rescorer.extractTerms` has been removed, and replaced with
a new method on `RescoreContext` that returns any queries used by the rescore implementation.
The delegating IndexSearcher then collects term contexts and statistics in the same way described
above for each Query.
This commit changes the `TransportVerifyShardBeforeCloseAction` so that it
always forces the flush of the shard. It seems that #37961 is not sufficient to
ensure that the translog and the Lucene commit share the exact same max
seq no and global checkpoint information in case of one or more noop
operations have been made.
The `BulkWithUpdatesIT.testThatMissingIndexDoesNotAbortFullBulkRequest`
and `FrozenIndexTests.testFreezeEmptyIndexWithTranslogOps` test this trivial
situation and they both fail 1 on 10 executions.
Relates to #33888
In #38333 and #38350 we moved away from the `discovery.zen` settings namespace
since these settings have an effect even though Zen Discovery itself is being
phased out. This change aligns the documentation and the names of related
classes and methods with the newly-introduced naming conventions.
We have had various reports of problems caused by the maxRetryTimeout
setting in the low-level REST client. Such setting was initially added
in the attempts to not have requests go through retries if the request
already took longer than the provided timeout.
The implementation was problematic though as such timeout would also
expire in the first request attempt (see #31834), would leave the
request executing after expiration causing memory leaks (see #33342),
and would not take into account the http client internal queuing (see #25951).
Given all these issues, it seems that this custom timeout mechanism
gives little benefits while causing a lot of harm. We should rather rely
on connect and socket timeout exposed by the underlying http client
and accept that a request can overall take longer than the configured
timeout, which is the case even with a single retry anyways.
This commit removes the `maxRetryTimeout` setting and all of its usages.
For some users, the built in authorization mechanism does not fit their
needs and no feature that we offer would allow them to control the
authorization process to meet their needs. In order to support this,
a concept of an AuthorizationEngine is being introduced, which can be
provided using the security extension mechanism.
An AuthorizationEngine is responsible for making the authorization
decisions about a request. The engine is responsible for knowing how to
authorize and can be backed by whatever mechanism a user wants. The
default mechanism is one backed by roles to provide the authorization
decisions. The AuthorizationEngine will be called by the
AuthorizationService, which handles more of the internal workings that
apply in general to authorization within Elasticsearch.
In order to support external authorization services that would back an
authorization engine, the entire authorization process has become
asynchronous, which also includes all calls to the AuthorizationEngine.
The use of roles also leaked out of the AuthorizationService in our
existing code that is not specifically related to roles so this also
needed to be addressed. RequestInterceptor instances sometimes used a
role to ensure a user was not attempting to escalate their privileges.
Addressing this leakage of roles meant that the RequestInterceptor
execution needed to move within the AuthorizationService and that
AuthorizationEngines needed to support detection of whether a user has
more privileges on a name than another. The second area where roles
leaked to the user is in the handling of a few privilege APIs that
could be used to retrieve the user's privileges or ask if a user has
privileges to perform an action. To remove the leakage of roles from
these actions, the AuthorizationService and AuthorizationEngine gained
methods that enabled an AuthorizationEngine to return the response for
these APIs.
Ultimately this feature is the work included in:
#37785#37495#37328#36245#38137#38219Closes#32435
Renames the following settings to remove the mention of `zen` in their names:
- `discovery.zen.hosts_provider` -> `discovery.seed_providers`
- `discovery.zen.ping.unicast.concurrent_connects` -> `discovery.seed_resolver.max_concurrent_resolvers`
- `discovery.zen.ping.unicast.hosts.resolve_timeout` -> `discovery.seed_resolver.timeout`
- `discovery.zen.ping.unicast.hosts` -> `discovery.seed_addresses`
Today the following settings in the `discovery.zen` namespace are still used:
- `discovery.zen.no_master_block`
- `discovery.zen.hosts_provider`
- `discovery.zen.ping.unicast.concurrent_connects`
- `discovery.zen.ping.unicast.hosts.resolve_timeout`
- `discovery.zen.ping.unicast.hosts`
This commit deprecates all other settings in this namespace so that they can be
removed in the next major version.
This PR removes the temporary change we made to the yml test harness in #37285
to automatically set `include_type_name` to `true` in index creation requests
if it's not already specified. This is possible now that the vast majority of
index creation requests were updated to be typeless in #37611. A few additional
tests also needed updating here.
Additionally, this PR updates the test harness to set `include_type_name` to
`false` in index creation requests when communicating with 6.x nodes. This
mirrors the logic added in #37611 to allow for typeless document write requests
in test set-up code. With this update in place, we can remove many references
to `include_type_name: false` from the yml tests.
The apache commons http client implementations recently released
versions that solve TLS compatibility issues with the new TLS engine
that supports TLSv1.3 with JDK 11. This change updates our code to
use these versions since JDK 11 is a supported JDK and we should
allow the use of TLSv1.3.
Today we pass `discovery.zen.minimum_master_nodes` to nodes started up in
tests, but for 7.x nodes this setting is not required as it has no effect.
This commit removes this setting so that nodes are started with more realistic
configurations, and deprecates it.
This PR attempts to remove all typed calls from our YAML REST tests. The PR adds include_type_name: false to create index requests that use a mapping and also to put mapping requests. It also removes _type from index requests where they haven't already been removed. The PR ignores tests named *_with_types.yml since this are specifically testing typed API behaviour.
The change also includes changing the test harness to add the type _doc to index, update, get and bulk requests that do not specify the document type when the test is running against a mixed 7.x/6.x cluster.
* Make repository settings override static settings
* Cache clients according to settings
* Introduce custom implementations for the AWS credentials here to be able to use them as part of a hash key
This commit moves the aggregation and mapping code from joda time to
java time. This includes field mappers, root object mappers, aggregations with date
histograms, query builders and a lot of changes within tests.
The cut-over to java time is a requirement so that we can support nanoseconds
properly in a future field mapper.
Relates #27330
This commit introduces a NetworkMessage class. This class has two
subclasses - InboundMessage and OutboundMessage. These messages can
be serialized and deserialized independent of the transport. This allows
more granular testing. Additionally, the serialization mechanism is now
a simple Supplier. This builds the framework to eventually move the
serialization of transport messages to the network thread. This is the
one serialization component that is not currently performed on the
network thread (transport deserialization and http serialization and
deserialization are all on the network thread).
Currently we create dedicated network threads for both the http and
transport implementations. Since these these threads should never
perform blocking operations, these threads could be shared. This commit
modifies the nio-transport to have 0 http workers be default. If the
default configs are used, this will cause the http transport to be run
on the transport worker threads. The http worker setting will still exist
in case the user would like to configure dedicated workers. Additionally,
this commmit deletes dedicated acceptor threads. We have never had these
for the netty transport and they can be added back if a need is
determined in the future.
The integ tests currently use the raw zip project name as the
distribution type. This commit simplifies this specification to be
"default" or "oss". Whether zip or tar is used should be an internal
implementation detail of the integ test setup, which can (in the future)
be platform specific.
This change adds a way to customize how phrase prefix queries should be created
on field types. The match phrase prefix query is exposed in field types in order
to allow optimizations based on the options set on the field.
For instance the text field uses the configured prefix field (if available) to
build a span near that mixes the original field and the prefix field on the last
position.
This change also contains a small refactoring of the match/multi_match query that
simplifies the interactions between the builders.
Closes#31921
The AbstracLifecycleComponent used to extend AbstractComponent, so it had to pass settings to the constractor of its supper class.
It no longer extends the AbstractComponent so there is no need for this constructor
There is also no need for AbstracLifecycleComponent subclasses to have Settings in their constructors if they were only passing it over to super constructor.
This is part 1. which will be backported to 6.x with a migration guide/deprecation log.
part 2 will have this constructor removed in 7
relates #35560
relates #34488
We recently migrated suggestions to `Writeable`. That allows us to also
clean up empty constructors and methods that called them as they are no
longer needed. They are replaced by constructors that accept a
`StreamInput` instance.
Today, a setting can declare that its validity depends on the values of other
related settings. However, the validity of a setting is not always checked
against the correct values of its dependent settings because those settings'
correct values may not be available when the validator runs.
This commit separates the validation of a settings updates into two phases,
with separate methods on the `Setting.Validator` interface. In the first phase
the setting's validity is checked in isolation, and in the second phase it is
checked again against the values of its related settings. Most settings only
use the first phase, and only the few settings with dependencies make use of
the second phase.
* There is no point in hsyncing after every individual write since there is only value in completely written blobs for restores, this is ensures by the `SYNC` flag already and there is no need for separately invoking `hsync` after individual writes
Today the routing of a SourceToParse is assigned in a separate step
after the object is created. We can easily forget to set the routing.
With this commit, the routing must be provided in the constructor of
SourceToParse.
Relates #36921
Today the default databases bundled with ingest-geoip are treated as
configuration files that we unbundle into the Elasticsearch
configuration directory. This can cause problems for users using our
Docker images if they bind mount over the configuration
directory. Additionally, it creates complexity when trying to convert
ingest-geoip to a module. This commit moves these databases out of the
configuration directory and instead loads from the plugin
directory. Further, custom databases can still be put into the
configuration directory.
Today we try to load the ingest-geoip databases lazily. Currently they
are loaded as soon as any pipeline that uses an ingest-geoip processor
is created. This is not lazy enough. For example, we could only load the
databases the first time that they are actually used. This would ensure
that we load the minimal set of data to support an in-use pipeline
(instead of *all* of the data). This can come up in a couple of
ways. One is when a subset of the database is used (e.g., the city
database versus the country database versus the ASN database). Another
is when the plugins are installed on non-ingest nodes (e.g., master-only
nodes); we would never use the database in this case yet they are
currently being loaded occupying ~60 MB of the heap. This commit makes
the ingest-geoip databases as lazy as possible.
Co-authored-by: Martijn van Groningen <martijn.v.groningen@gmail.com>
Christoph Büscher
Jason Tedor
Mark Vieira
Armin Braun
Simon Willnauer
Jim Ferenczi
Jim Ferenczi
Jay Modi
Alan Woodward
Alpar Torok
Yannick Welsch
David Turner
Tim Brooks
Alpar Torok
clement-tourriere
Martijn van Groningen
Andy Bristol
Henning Andersen
Ryan Ernst
David Emanuel Buchmann
markharwood
Alan Woodward
Tanguy Leroux
Luca Cavanna
Julie Tibshirani
Colin Goodheart-Smithe
Alexander Reelsen
Adrien Grand
Przemyslaw Gomulka
Daniel Mitterdorfer
Christophe Bismuth
Nick Knize
Nhat Nguyen