The date_histogram internally converts obsolete timezones (such as
"Canada/Mountain") into their modern equivalent ("America/Edmonton").
But rollup just stored the TZ as provided by the user.
When checking the TZ for query validation we used a string comparison,
which would fail due to the date_histo's upgrading behavior.
Instead, we should convert both to a TimeZone object and check if their
rules are compatible.
The `ignore_malformed` option currently works on numeric fields only when the
bad value isn't a string value but not if it is a boolean. In this case we get a
parsing error from the xContent parser which we need to catch in addition to the
field mapper.
Closes#11498
Today the `?preference=custom_string_value` search preference will only change
its choice of a shard copy if something changes the `IndexShardRoutingTable`
for that specific shard. Users can use this behaviour to route searches to a
consistent set of shard copies, which means they can reliably hit copies with
hot caches, and use the other copies only for redundancy in case of failure.
However we do not assert this property anywhere, so we might break it in
future.
This commit adds a test that shows that searches are routed consistently even
if other indices are created/rebalanced/deleted.
Relates https://discuss.elastic.co/t/176598, #41115, #26791
Today we always trim unsafe commits (whose max_seq_no >= global
checkpoint) before starting a read-write or read-only engine. This is
mandatory for read-write engines because they must start with the safe
commit. This is also fine for read-only engines since most of the cases
we should have exactly one commit after closing an index (trimming is a
noop). However, this is dangerous for following indices which might have
more than one commits when they are being closed.
With this change, we move the trimming logic to the ctor of InternalEngine
so we won't trim anything if we are going to open a read-only engine.
Currently enabling profiling disables top-hits optimizations, which is
unfortunate: it would be nice to be able to notice the difference in method
counts and timings depending on whether total hit counts are requested.
`Node#close` is pretty hard to rely on today:
- it might swallow exceptions
- it waits for 10 seconds for threads to terminate but doesn't signal anything
if threads are still not terminated after 10 seconds
This commit makes `IOException`s propagated and splits `Node#close` into
`Node#close` and `Node#awaitClose` so that the decision what to do if a node
takes too long to close can be done on top of `Node#close`.
It also adds synchronization to lifecycle transitions to make them atomic. I
don't think it is a source of problems today, but it makes things easier to
reason about.
Today we check if an index has broken settings when checking if an index
needs to be upgraded. However, it can be the case that an index setting
became broken even if an index is already upgraded to the current
version if the user removed a plugin (or downgraded from the default
distribution to the non-default distribution) while on the same version
of Elasticsearch. In this case, some registered settings would go
missing and the index would now be broken. Yet, we miss this check and
instead of archiving the settings, the index becomes unassigned due to
the missing settings. This commit addresses this by checking for broken
settings whether or not the index is upgraded.
One of the two #getCorrections methods is only used in tests, so we can move
it and any of the required helper methods to that test. Also reducing the
visibility of several methods to package private since the class isn't used
elsewhere outside the package.
Today we erroneously look for a node setting called `readonly` when deciding
whether or not to create a missing directory in a filesystem repository. This
change fixes this by using the repository setting instead.
Closes#41009
Relates #26909
In `TransportRolloverAction` before doing rollover we resolve
source index name (write index) from the alias in the rollover request.
Before evaluating the conditions and executing rollover action, we
retrieve stats, but to do so we used the source index name
resolved from the alias instead of alias from the index.
This fails when the user is assigned a role with index privilege on the
alias instead of the concrete index. This commit fixes this by using
the alias from the request.
After this change, verified that when we retrieve all the stats (including write + read indexes)
we are considering only source index.
Closes#40771
The unified highlighter returns the first sentence of the text when number_of_fragments
is set to 0 (full highlighting). This is a legacy of the removed postings highlighter
that was based on sentence break only. This commit changes this behavior in order
to respect the provided no_match_size value when number_of_fragments is set to 0.
This means that the behavior will be consistent for any value of the number_of_fragments option.
Closes#41066
* 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
Today the blended term query detects if a term exists in a field by looking at the term statistics in the index.
However the value to indicate that a term has no occurence in a field have changed in Lucene. A non-existing term now returns
a doc and total term frequency of 0. Because of this disrepancy the blended term query picks 0 as the minimum frequency for a term
even if other fields have documents for this terms. This confuses the term queries that the blending creates since some of them
contain a custom state that indicates a frequency of 0 even though the term has some occurence in the field. For these terms an exception
is thrown because the term query always checks that the term state's frequency is greater than 0 if there are documents associate to it.
This change fixes this bug by ignoring terms with a doc freq of 0 when the blended term query picks the minimum term frequency among the
requested fields.
Closes#41118
`TransportReplicationAction.AsyncPrimaryAction#createReplicatedOperation`
exists so it can be overridden in tests. This commit re-works these tests to
use a real `ReplicationOperation` and inlines the now-unnecessary method.
Relates #40706.
Today we fail to join a Zen2 cluster if the cluster UUID does not match our
own, but we do not perform the same validation when joining a Zen1 cluster.
This means that a Zen2 node will pass join validation and be added to a Zen1
cluster but will reject all cluster states from the master.
Relates #37775
Currently we throw an error when a range querys minimum value exceeds the
maximum value due to the fact that they are neighbouring values and both upper
and lower value are excluded from the interval.
Since this is a condition that the user usually doesn't specify conciously (at
least in the case of float and double values its difficult to see which values
are adjacent) we should ignore those "wrong" intervals and create a
MatchNoDocsQuery in those cases.
We should still throw errors with an actionable message if the user specifies
the query interval in a way that min value > max value. This PR adds those
checks and tests for those cases.
Closes#40937
This is related to #36652. We intend to deprecate a number of transport
settings in 7.x and remove them in 8.0. This commit removes the string
usages of these settings.
Pipelines require single-valued agg or a numeric to be returned.
If they don't get that, they throw an exception. Unfortunately, this
exception text is very confusing to users because it usually arises
from pathing "through" multiple terms aggs. The final target is a numeric,
but it's the intermediary aggs that cause the problem.
This commit adds the current agg name to the exception message
so the user knows which "level" is the issue.
Full text queries ignore unmapped fields since https://github.com/elastic/elasticsearch/issues/41022
even if all fields in the query are unmapped.
This change makes sure that we ignore unmapped fields only if they are mixed
with mapped fields and returns a MatchNoDocsQuery otherwise.
Closes#41022
This change adds either ToXContentObject or ToXContentFragment to classes
directly implementing ToXContent currently. This helps in reasoning about
whether those implementations output full xcontent object or just fragments.
Relates to #16347
When a polygon contains a self-intersection due to have twice the same point in no-consecutive position, the polygon builder tries to split the polygon. During the split one of the polygons become invalid as it is not closed and an error is thrown which is not related to the real issue.
We detect this situation now and throw a more meaningful error.
Today a new replica of a closed index does not have a safe commit
invariant when its engine is opened because we won't initialize the
global checkpoint on a recovering replica until the finalize step. With
this change, we can achieve that property by creating a new translog
with the global checkpoint from the primary at the end of phase 1.
This helps avoid memory issues when computing deep sub-aggregations. Because it
should be rare to use sub-aggregations with significant terms, we opted to always
choose breadth first as opposed to exposing a `collect_mode` option.
Closes#28652.
Since #40249, we always reinitialize max_seq_no_of_updates to max_seq_no
when a promoting primary restores history regardless of whether it did
rollback previously or not.
Closes#40929
This commit removes the settings member variable from Node.
This member made it confusing which settings should actually be looked
at. Now all settings are accessed through the final environment.
A small refactoring that removes the primaryTerm field from ReplicasProxy and
instead passes it directly in to the methods that need it. Relates #40706.
This is a dependency of #39504
Motivation:
By refactoring `TransportShardBulkAction#shardOperationOnPrimary` to async, we enable using `DeterministicTaskQueue` based tests to run indexing operations. This was previously impossible since we were blocking on the `write` thread until the `update` thread finished the mapping update.
With this change, the mapping update will trigger a new task in the `write` queue instead.
This change significantly enhances the amount of coverage we get from `SnapshotResiliencyTests` (and other potential future tests) when it comes to tracking down concurrency issues with distributed state machines.
The logical change is effectively all in `TransportShardBulkAction`, the rest of the changes is then simply mechanically moving the caller code and tests to being async and passing the `ActionListener` down.
Since the move to async would've added more parameters to the `private static` steps in this logic, I decided to inline and dry up (between delete and update) the logic as much as I could instead of passing the listener + wait-consumer down through all of them.
* 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
It could be that we try to shutdown the executor pool before all the
listeners have been invoked. It can happen that one was not invoked if
it timed out and was in the process of being notified that it timed out
on the executor. If we do this shutdown then, a listener will be met
with rejected execution exception. To address this, we first wait until
all listeners have been notified (or timed out) before proceeding with
shutting down the executor.
Relates #40970
Added documentation for node repurpose tool and included documentation on how to repurpose nodes safely. Adjusted order of tools in `elasticsearch-node` tool since the repurpose tool is most likely to be used.
Co-Authored-By: David Turner <david.turner@elastic.co>
Today if `cluster.routing.rebalance.enable: none` then rebalancing is disabled,
but we still execute `balanceByWeights()` and perform some rather expensive
calculations before discovering that we cannot rebalance any shards. In a large
cluster this can make cluster state updates occur rather slowly. With this
change we check earlier whether rebalancing is globally disabled and, if so,
avoid the rebalancing process entirely.
Relates #40942 which was reverted because of egregiously faulty tests.
The number of user data attributes of an index commit has increased
from 6 to 8, but we forgot to adjust. This change increases the initial
size of that map to avoid resizing.
Reducing some methods scope and marking them as static where possible. Removing
"alias" support from AnalysisRegistry#produceAnalyze and changing that method to
return a NamedAnalyzer instead of having a side effect on the analyzer map passed in.
Also, CustomAnalyzerProvider doesn't seem to need the `environment` field.
Today if `cluster.routing.rebalance.enable: none` then rebalancing is disabled,
but we still execute `balanceByWeights()` and perform some rather expensive
calculations before discovering that we cannot rebalance any shards. In a large
cluster this can make cluster state updates occur rather slowly. With this
change we check earlier whether rebalancing is globally disabled and, if so,
avoid the rebalancing process entirely.
The phrase suggesters have an option to remove terms that have
a frequency lower than a provided min_doc_freq. However this value is
overwritten by the frequency of the original term in the popular mode.
This change ensures that we keep the maximum value between the provided
min_doc_value and the original term frequency as a threshold to select
candidates.
Fixes#16764
Today we are strict when parsing build flavor and types off the
wire. This means that if a later version introduces a new build flavor
or type, an older version would not be able to parse what that new
version is sending. For a practical example of this, we recently added
the build type "docker", and this means that in a rolling upgrade
scenario older nodes would not be able to understand the build type that
the newer node is sending. This breaks clusters and is bad. We do not
normally think of adding a new enumeration value as being a
serialization breaking change, it is just not a lesson that we have
learned before. We should be lenient here though, so that we can add
future changes without running the risk of breaking ourselves
horribly. It is either that, or we have super-strict testing
infrastructure here yet still I fear the possibility of mistakes. This
commit changes the parsing of build flavor and build type so that we are
still strict at startup, yet we are lenient with values coming across
the wire. This will help avoid us breaking rolling upgrades, or clients
that are on an older version.
If the transport service is stopped, likely because we are shutting
down, and a retention lease background sync fires the logs will display
a warn message and stacktrace. Yet, this situaton is harmless and can
happen as a normal course of business when shutting down. This commit
suppresses the log messages in this case.
The `getIndexShard()` and `sendReplicaRequest()` methods in
TransportReplicationAction are effectively only used to customise some
behaviour in tests. However there are other ways to do this that do not cause
such an obstacle to separating the TransportReplicationAction into its two
halves (see #40706).
This commit removes these customisation points and injects the test-only
behaviour using other techniques.
Many gradle projects specifically use the -try exclude flag, because
there are many cases where auto-closeable resource ignore is never
referenced in body of corresponding try statement. Suppressing this
warning specifically in each case that it happens using
`@SuppressWarnings("try")` would be very verbose.
This change removes `-try` from any gradle project and adds it to the
build plugin. Also this change removes exclude flags from gradle projects
that is already specified in build plugin (for example -deprecation).
Relates to #40366
Primary-replica resync in a mixed-cluster between 6.x and 5.6 can send
operations without sequence number to a replica which already processed
operations with sequence number. This leads to the failure of that
replica for we trip the sequence number assertion when writing resync
operations without sequence number to translog.
This change rejects an illegal combination of flush parameters where
force is true, but wait_if_ongoing is false. This combination is trappy
and should be forbidden.
Closes#36342
If there's an ongoing flush triggered by the translog flush threshold,
we may fail to execute a flush because waitIfOngoing is false by
default.
Relates to #36342
If a refresh, which is scheduled by the setting change, executes after
the index-2 operation and win the refresh race (i.e., maybeRefresh) with
the scheduledRefresh that we are going to check, then the latter will
return false.
Closes#39565
Relates #39462
PR #40387
A user reported that the same query that takes ~900ms when querying an index
pattern only takes ~50ms when only querying indices that have matches. The
query is a date range query and we confirmed that the `can_match` phase works
as expected. I was able to reproduce this issue locally with a single node: with
900 1-shard indices, a query to an index pattern that matches all indices runs
in ~90ms while a query to the only index that has matches runs in 0-1ms.
This ended up not being related to the `can_match` phase but to the cost of
resolving aliases when querying an index pattern that matches lots of indices.
In that case, we first resolve the index pattern to a list of concrete indices
and then for each concrete index, we check whether it was matched through an
alias, meaning we might have to apply alias filters. Unfortunately this second
per-index operation runs in linear time with the number of matched concrete
indices, which means that alias resolution runs in O(num_indices^2) overall.
So queries get exponentially slower as an index pattern matches more indices.
I reorganized alias resolution into a one-step operation that runs in linear
time with the number of matches indices, and then a per-index operation that
runs in linear time with the number of aliases of this index. This makes alias
resolution run is O(num_indices * num_aliases_per_index) overall instead. When
testing the scenario described above, the `took` went down from ~90ms to ~10ms.
It is still more than the 0-1ms latency that one gets when only querying the
single index that has data, but still much better than what we had before.
Closes#40248
If a field `field_name` was missing in a document,
doc['field_name'].get(0) incorrectly retrieved
a value of the previously accessed document.
This happened because `get(int index)` function
was just accessing `values[index]` without
checking the number of values - `count`.
This PR fixes this.
There were some test failures caused by the background retention lease sync running on a relocated
primary. This commit fixes the situation that triggered the assertion and reactivates the failing test.
Closes#40731
The Eclipse compiler (4.10, Photon) cannot build this test because it cannot
correctly infer the type arguments of the functions. Explicitely adding them
helps in this case.
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
It is important that resync actions are not rejected on the primary even if its
`write` threadpool is overloaded. Today we do this by exposing
`registerRequestHandlers` to subclasses and overriding it in
`TransportResyncReplicationAction`. This isn't ideal because it obscures the
difference between this action and other replication actions, and also might
allow subclasses to try and use some state before they are properly
initialised. This change replaces this override with a constructor parameter to
solve these issues.
Relates #40706
This commit deprecates versions of Java prior to Java 11. This commit
will cause a warning to be printed to standard error when any command
line tool is invoked, or when Elasticsearch is started. Additionally, we
log a deprecation message when Elasticsearch is started.
`TransportReplicationAction` is a rather complex beast, and some of its
concrete implementations do not need all of its features. More specifically, it
(a) chases a primary around the cluster until it manages to pin it down and
then (b) executes an action on that primary and all its replicas. There are
some actions that are coordinated by the primary itself, meaning that there is
no need for the chase-the-primary phases, and in the case of peer recovery
retention leases and primary/replica resync it is important to bypass these
first phases.
This commit is a step towards separating the `TransportReplicationAction` into
these two parts. It is a mostly mechanical sequence of steps to remove some
abstractions that are no longer in use.
Completion and DocStats are pulled from internal readers
instead of external since #33835 and #33847 which doesn't require
us to refresh after a stats call since refreshes will happen internally
anyhow and that will cause updated stats on ongoing indexing.
In 6.7.0 (#39378) we added a build type of DOCKER for the docker images, but
unfortunately earlier versions do not understand this and will reject any
transport messages that mention this build type.
This commit fixes this by reporting TAR instead of DOCKER when talking to older
nodes.
Relates (but does not fix) #40511
Relates #39378
In order to remain compatible with the existing joda based
implementation the parsing of milliseconds should support parsing single
digits instead of relying on three, even with strict formats.
This adds a few tests to duel against the existing joda based
implementation in order to ensure the parsing behaviour is the same.
Closes#40403
Currently, if Manifest write is unsuccessful (i.e. WriteStateException
is thrown) we perform cleanup of newly created metadata files.
However, this is wrong.
Consider the following sequence (caught by CI here
https://github.com/elastic/elasticsearch/issues/39077):
- cluster global data is written **successful**
- the associated manifest write **fails** (during the fsync, ie files
have been written)
- deleting (revert) the manifest files, **fails**, metadata is
therefore persisted
- deleting (revert) the cluster global data is **successful**
In this case, when trying to load metadata (after node restart
because of dirty WriteStateException), the following exception will
happen
```
java.io.IOException: failed to find global metadata [generation: 0]
```
because the manifest file is referencing missing global metadata file.
This commit checks if thrown WriteStateException is dirty and if its
we don't perform any cleanup, because new Manifest file might be
created, but its deletion has failed.
In the future, we might add more fine-grained check - perform the
clean up if WriteStateException is dirty, but Manifest deletion is
successful.
Closes https://github.com/elastic/elasticsearch/issues/39077
(cherry picked from commit 1fac56916bb3c4f3333c639e59188dbe743e385b)
On mapping updates the `text` field mapper does not update
the field types for the underlying prefix and phrase fields.
In practice this shouldn't be considered as a bug but we have
an assert in the code that check that field types in the mapper service
are identical to the ones present in field mappers.
When geo point parsing threw a parse exception, it did not consume
remaining tokens from the parser. This in turn meant that
indexing documents with malformed geo points into mappings with
ignore_malformed=true would fail in some cases, since DocumentParser
expects geo_point parsing to end on the END_OBJECT token.
Related to #17617
As part of #40177 we have added top-level pipeline aggs to
`InternalAggregations`. Given that `QuerySearchResult` holds an
`InternalAggregations` instance, there is no need to keep on setting
top-level pipeline aggs separately. Top-level pipeline aggs can then
always be transported through `InternalAggregations`. Such change is
made in a backwards compatible manner.
IOException are never thrown in any of the existing pipeline aggregation
builders. Removing the throws IOException from the create method allows
to remove it also from a couple of other methods which ends up simplifying
AggregationPhase (one less catch).
The cat recovery API is incredibly useful. Yet it is missing the start
and stop time as an option from the output. This commit adds these as
options to the cat recovery API. We elect to make these not visible by
default to avoid breaking the output that users might rely on.
To make script_score query to have the same features
as function_score query, we need to add randomScore
function.
This function produces different
random scores on different index shards.
It is also able to produce random scores
based on the internal Lucene Document Ids.
Today if you try and insert a very large number like `1e9999999` into a long
field we first construct this number as a `BigDecimal`, convert this to a
`BigInteger` and then reject it because it is out of range. Unfortunately
making such a large `BigInteger` is rather expensive.
We can avoid this expense by performing a (weaker) range check on the
`BigDecimal` representation of incoming `long`s too.
Relates #26137Closes#40323
In #33062 we introduced the `cluster.remote.*.proxy` setting for proxied
connections to remote clusters, but left it deliberately undocumented since it
needed followup work so that it could work with SNI. However, since #32517 is
now closed we can add this documentation and remove the comment about its lack
of documentation.
This commit fixes an edge case in tests where search hits are empty
after the merge but some shards returned hits. This can happen if
the total number of merged hits is less than the provided `from`.
Closes#40553
It initially mentioned the type in the exception because the type used to be
required to uniquely identify a document. This is not necessary anymore given
that indices have at most one type.
`Index` interns its name and uuid. My guess is that the main goal is to avoid
having duplicate strings in the representation of the cluster state. However
I doubt it helps much given that we have many other objects in the cluster state
that we don't try to reuse, and interning has some cost. When looking into
#40263 my profiler pointed to string interning because of the `Index` object
that is created in `QueryShardContext` as one of the bottlenecks of the
`can_match` phase.
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
Replicated closed indices can't be indexed into or searched, and therefore don't need a shard with
full indexing and search capabilities allocated. We can save on a lot of heap memory for those
indices by not allocating a mapper service and caching infrastructure (which preallocates a constant
amount per instance). Before this change, a 1GB ES instance could host 250 replicated closed
metricbeat indices (each index with one shard). After this change, the same instance can host 7300
replicated closed metricbeat instances (not that this would be a recommended configuration). Most
of the remaining memory is in the cluster state and the IndexSettings object.
Switches "discovery.type: single-node" from using a separate implementation for single-node discovery to using the existing standard discovery implementation, with two small adaptions:
- auto-bootstrapping, but requiring initial_master_nodes not to be set.
- not actively pinging other nodes using the Peerfinder
- not allowing other nodes to join its single-node cluster (if they have e.g. been set up using regular discovery and connect to the single-disco node).
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.
This test inadvertently asserts that the election occurs after a master failure
is clean. However, messy elections are a fact of life so we should not fail on
a messy election.
This change moves this test away from an `AbstractDisruptionTestCase` since it
does not need the fault detector to be so enthusiastic, and weakens the
assertions to merely say that we ignore states published by the old master
without saying anything about the cleanliness of the election.
Closes#36556
Currently the TransportMessageListener is applied and used in the
Transport class. However, local requests and responses never make it to
this class. This PR moves the listener add/remove methods to the
TransportService. After this change the Transport can only have one
listener set with it. This one listener is the TransportService, which
will then propogate the events to the external listeners.
Additionally this commit back ports #40237
Remove Tracer from MockTransportService
Currently the TransportMessageListener is applied and used in the
Transport class. However, local requests and responses never make it to
this class. This PR moves the listener add/remove methods to the
TransportService. After this change the Transport can only have one
listener set with it. This one listener is the TransportService, which
will then propogate the events to the external listeners.
Java-time fails parsing composite patterns when first pattern matches only the prefix of the input. It expects pattern in longest to shortest order. Because of this constructing just one DateTimeFormatter with appendOptional is not sufficient. Parsers have to be iterated and if the parsing fails, the next one in order should be used. In order to not degrade performance parsing should not be throw exceptions on failure. Format.parseObject was used as it only returns null when parsing failed and allows to check if full input was read.
closes#39916
backport #40100
Zachary Tong
Christoph Büscher
David Turner
Nhat Nguyen
Adrien Grand
Jason Tedor
Yogesh Gaikwad
Jim Ferenczi
Armin Braun
Tim Brooks
Ignacio Vera
Antonio Matarrese
Ryan Ernst
Zachary Tong
Dimitrios Liappis
Julie Tibshirani
Mark Vieira
Henning Andersen
Colin Goodheart-Smithe
Martijn van Groningen
Mayya Sharipova
Yannick Welsch
Simon Willnauer
Alexander Reelsen
Andrey Ershov
Luca Cavanna
Oghenovo Usiwoma
jimczi
Andy Bristol
Benjamin Trent
Przemyslaw Gomulka