* Kinesis: More robust default fetch settings.
1) Default recordsPerFetch and recordBufferSize based on available memory
rather than using hardcoded numbers. For this, we need an estimate
of record size. Use 10 KB for regular records and 1 MB for aggregated
records. With 1 GB heaps, 2 processors per task, and nonaggregated
records, recordBufferSize comes out to the same as the old
default (10000), and recordsPerFetch comes out slightly lower (1250
instead of 4000).
2) Default maxRecordsPerPoll based on whether records are aggregated
or not (100 if not aggregated, 1 if aggregated). Prior default was 100.
3) Default fetchThreads based on processors divided by task count on
Indexers, rather than overall processor count.
4) Additionally clean up the serialized JSON a bit by adding various
JsonInclude annotations.
* Updates for tests.
* Additional important verify.
This commit adds a new class `InputStats` to track the total bytes processed by a task.
The field `processedBytes` is published in task reports along with other row stats.
Major changes:
- Add class `InputStats` to track processed bytes
- Add method `InputSourceReader.read(InputStats)` to read input rows while counting bytes.
> Since we need to count the bytes, we could not just have a wrapper around `InputSourceReader` or `InputEntityReader` (the way `CountableInputSourceReader` does) because the `InputSourceReader` only deals with `InputRow`s and the byte information is already lost.
- Classic batch: Use the new `InputSourceReader.read(inputStats)` in `AbstractBatchIndexTask`
- Streaming: Increment `processedBytes` in `StreamChunkParser`. This does not use the new `InputSourceReader.read(inputStats)` method.
- Extend `InputStats` with `RowIngestionMeters` so that bytes can be exposed in task reports
Other changes:
- Update tests to verify the value of `processedBytes`
- Rename `MutableRowIngestionMeters` to `SimpleRowIngestionMeters` and remove duplicate class
- Replace `CacheTestSegmentCacheManager` with `NoopSegmentCacheManager`
- Refactor `KafkaIndexTaskTest` and `KinesisIndexTaskTest`
* Zero-copy local deep storage.
This is useful for local deep storage, since it reduces disk usage and
makes Historicals able to load segments instantaneously.
Two changes:
1) Introduce "druid.storage.zip" parameter for local storage, which defaults
to false. This changes default behavior from writing an index.zip to writing
a regular directory. This is safe to do even during a rolling update, because
the older code actually already handled unzipped directories being present
on local deep storage.
2) In LocalDataSegmentPuller and LocalDataSegmentPusher, use hard links
instead of copies when possible. (Generally this is possible when the
source and destination directory are on the same filesystem.)
Main changes:
1) Convert SeekableStreamIndexTaskClient to an interface, move old code
to SeekableStreamIndexTaskClientSyncImpl, and add new implementation
SeekableStreamIndexTaskClientAsyncImpl that uses ServiceClient.
2) Add "chatAsync" parameter to seekable stream supervisors that causes
the supervisor to use an async task client.
3) In SeekableStreamSupervisor.discoverTasks, adjust logic to avoid making
blocking RPC calls in workerExec threads.
4) In SeekableStreamSupervisor generally, switch from Futures.successfulAsList
to FutureUtils.coalesce, so we can better capture the errors that occurred
with contacting individual tasks.
Other, related changes:
1) Add ServiceRetryPolicy.retryNotAvailable, which controls whether
ServiceClient retries unavailable services. Useful since we do not
want to retry calls unavailable tasks within the service client. (The
supervisor does its own higher-level retries.)
2) Add FutureUtils.transformAsync, a more lambda friendly version of
Futures.transform(f, AsyncFunction).
3) Add FutureUtils.coalesce. Similar to Futures.successfulAsList, but
returns Either instead of using null on error.
4) Add JacksonUtils.readValue overloads for JavaType and TypeReference.
* Support for middle manager less druid, tasks launch as k8s jobs
* Fixing forking task runner test
* Test cleanup, dependency cleanup, intellij inspections cleanup
* Changes per PR review
Add configuration option to disable http/https proxy for the k8s client
Update the docs to provide more detail about sidecar support
* Removing un-needed log lines
* Small changes per PR review
* Upon task completion we callback to the overlord to update the status / locaiton, for slower k8s clusters, this reduces locking time significantly
* Merge conflict fix
* Fixing tests and docs
* update tiny-cluster.yaml
changed `enableTaskLevelLogPush` to `encapsulatedTask`
* Apply suggestions from code review
Co-authored-by: Abhishek Agarwal <1477457+abhishekagarwal87@users.noreply.github.com>
* Minor changes per PR request
* Cleanup, adding test to AbstractTask
* Add comment in peon.sh
* Bumping code coverage
* More tests to make code coverage happy
* Doh a duplicate dependnecy
* Integration test setup is weird for k8s, will do this in a different PR
* Reverting back all integration test changes, will do in anotbher PR
* use StringUtils.base64 instead of Base64
* Jdk is nasty, if i compress in jdk 11 in jdk 17 the decompressed result is different
Co-authored-by: Rahul Gidwani <r_gidwani@apple.com>
Co-authored-by: Abhishek Agarwal <1477457+abhishekagarwal87@users.noreply.github.com>
* Remove basePersistDirectory from tuning configs.
Since the removal of CliRealtime, it serves no purpose, since it is
always overridden in production using withBasePersistDirectory given
some subdirectory of the task work directory.
Removing this from the tuning config has a benefit beyond removing
no-longer-needed logic: it also avoids the side effect of empty
"druid-realtime-persist" directories getting created in the systemwide
temp directory.
* Test adjustments to appropriately set basePersistDirectory.
* Remove unused import.
* Fix RATC constructor.
Kinesis ingestion requires all shards to have at least 1 record at the required position in druid.
Even if this is satisified initially, resharding the stream can lead to empty intermediate shards. A significant delay in writing to newly created shards was also problematic.
Kinesis shard sequence numbers are big integers. Introduce two more custom sequence tokens UNREAD_TRIM_HORIZON and UNREAD_LATEST to indicate that a shard has not been read from and that it needs to be read from the start or the end respectively.
These values can be used to avoid the need to read at least one record to obtain a sequence number for ingesting a newly discovered shard.
If a record cannot be obtained immediately, use a marker to obtain the relevant shardIterator and use this shardIterator to obtain a valid sequence number. As long as a valid sequence number is not obtained, continue storing the token as the offset.
These tokens (UNREAD_TRIM_HORIZON and UNREAD_LATEST) are logically ordered to be earlier than any valid sequence number.
However, the ordering requires a few subtle changes to the existing mechanism for record sequence validation:
The sequence availability check ensures that the current offset is before the earliest available sequence in the shard. However, current token being an UNREAD token indicates that any sequence number in the shard is valid (despite the ordering)
Kinesis sequence numbers are inclusive i.e if current sequence == end sequence, there are more records left to read.
However, the equality check is exclusive when dealing with UNREAD tokens.
* Mid-level service client and updated high-level clients.
Our servers talk to each other over HTTP. We have a low-level HTTP
client (HttpClient) that is super-asynchronous and super-customizable
through its handlers. It's also proven to be quite robust: we use it
for Broker -> Historical communication over the wide variety of query
types and workloads we support.
But the low-level client has no facilities for service location or
retries, which means we have a variety of high-level clients that
implement these in their own ways. Some high-level clients do a better
job than others. This patch adds a mid-level ServiceClient that makes
it easier for high-level clients to be built correctly and harmoniously,
and migrates some of the high-level logic to use ServiceClients.
Main changes:
1) Add ServiceClient org.apache.druid.rpc package. That package also
contains supporting stuff like ServiceLocator and RetryPolicy
interfaces, and a DiscoveryServiceLocator based on
DruidNodeDiscoveryProvider.
2) Add high-level OverlordClient in org.apache.druid.rpc.indexing.
3) Indexing task client creator in TaskServiceClients. It uses
SpecificTaskServiceLocator to find the tasks. This improves on
ClientInfoTaskProvider by caching task locations for up to 30 seconds
across calls, reducing load on the Overlord.
4) Rework ParallelIndexSupervisorTaskClient to use a ServiceClient
instead of extending IndexTaskClient.
5) Rework RemoteTaskActionClient to use a ServiceClient instead of
DruidLeaderClient.
6) Rework LocalIntermediaryDataManager, TaskMonitor, and
ParallelIndexSupervisorTask. As a result, MiddleManager, Peon, and
Overlord no longer need IndexingServiceClient (which internally used
DruidLeaderClient).
There are some concrete benefits over the prior logic, namely:
- DruidLeaderClient does retries in its "go" method, but only retries
exactly 5 times, does not sleep between retries, and does not retry
retryable HTTP codes like 502, 503, 504. (It only retries IOExceptions.)
ServiceClient handles retries in a more reasonable way.
- DruidLeaderClient's methods are all synchronous, whereas ServiceClient
methods are asynchronous. This is used in one place so far: the
SpecificTaskServiceLocator, so we don't need to block a thread trying
to locate a task. It can be used in other places in the future.
- HttpIndexingServiceClient does not properly handle all server errors.
In some cases, it tries to parse a server error as a successful
response (for example: in getTaskStatus).
- IndexTaskClient currently makes an Overlord call on every task-to-task
HTTP request, as a way to find where the target task is. ServiceClient,
through SpecificTaskServiceLocator, caches these target locations
for a period of time.
* Style adjustments.
* For the coverage.
* Adjustments.
* Better behaviors.
* Fixes.
listShards API was used to get all the shards for kinesis ingestion to improve its resiliency as part of #12161.
However, this may require additional permissions in the IAM policy where the stream is present. (Please refer to: https://docs.aws.amazon.com/kinesis/latest/APIReference/API_ListShards.html).
A dynamic configuration useListShards has been added to KinesisSupervisorTuningConfig to control the usage of this API and prevent issues upon upgrade. It can be safely turned on (and is recommended when using kinesis ingestion) by setting this configuration to true.
* Store null columns in the segments
* fix test
* remove NullNumericColumn and unused dependency
* fix compile failure
* use guava instead of apache commons
* split new tests
* unused imports
* address comments
This PR aims to make the ParseExceptions in Druid more informative, by adding additional information (metadata) to the ParseException, which can contain additional information about the exception. For example - the path of the file generating the issue, the line number (where it can be easily fetched - like CsvReader)
Following changes are addressed in this PR:
A new class CloseableIteratorWithMetadata has been created which is like CloseableIterator but also has a metadata method that returns a context Map<String, Object> about the current element returned by next().
IntermediateRowParsingReader#read() now attaches the InputEntity and the "record number" which created the exception (while parsing them), and IntermediateRowParsingReader#sample attaches the InputEntity (but not the "record number").
TextReader (and its subclasses), which is a specific implementation of the IntermediateRowParsingReader also include the line number which caused the generation of the error.
This will also help in triaging the issues when InputSourceReader generates ParseException because it can point to the specific InputEntity which caused the exception (while trying to read it).
Problem:
- When a kinesis stream is resharded, the original shards are closed.
Any intermediate shard created in the process is eventually closed as well.
- If a shard is closed before any record is put into it, it can be safely ignored for ingestion.
- It is expensive to determine if a closed shard is empty, since it requires a call to the Kinesis cluster.
Changes:
- Maintain a cache of closed empty and closed non-empty shards in `KinesisSupervisor`
- Add config `skipIngorableShards` to `KinesisSupervisorTuningConfig`
- The caches are used and updated only when `skipIgnorableShards = true`
Makes kinesis ingestion resilient to `LimitExceededException` caused by resharding.
Replace `describeStream` with `listShards` (recommended) to get shard related info.
`describeStream` has a limit (100) to the number of shards returned per call and a low default TPS limit of 10.
`listShards` returns the info for at most 1000 shards and has a higher TPS limit of 100 as well.
Key changed/added classes in this PR
* `KinesisRecordSupplier`
* `KinesisAdminClient`
This fixes a bug that causes TaskClient in overlord to continuously retry to pause tasks. This can happen when a task is not responding to the pause command. Ideally, in such a case when the task is unresponsive, the overlord would have given up after a few retries and would have killed the task. However, due to this bug, retries go on forever.
* Enhancements to IndexTaskClient.
1) Ability to use handlers other than StringFullResponseHandler. This
functionality is not used in production code yet, but is useful
because it will allow tasks to communicate with each other in
non-string-based formats and in streaming fashion. In the future,
we'll be able to use this to make task-to-task communication
more efficient.
2) Truncate server errors at 1KB, so long errors do not pollute logs.
3) Change error log level for retryable errors from WARN to INFO. (The
final error is still WARN.)
4) Harmonize log and exception messages to have a more consistent format.
* Additional tests and improvements.
* Make persists concurrent with ingestion
* Remove semaphore but keep concurrent persists (with add) and add push in the backround as well
* Go back to documented default persists (zero)
* Move to debug
* Remove unnecessary Atomics
* Comments on synchronization (or not) for sinks & sinkMetadata
* Some cleanup for unit tests but they still need further work
* Shutdown & wait for persists and push on close
* Provide support for three existing batch appenderators using batchProcessingMode flag
* Fix reference to wrong appenderator
* Fix doc typos
* Add BatchAppenderators class test coverage
* Add log message to batchProcessingMode final value, fix typo in enum name
* Another typo and minor fix to log message
* LEGACY->OPEN_SEGMENTS, Edit docs
* Minor update legacy->open segments log message
* More code comments, mostly small adjustments to naming etc
* fix spelling
* Exclude BtachAppenderators from Jacoco since it is fully tested but Jacoco still refuses to ack coverage
* Coverage for Appenderators & BatchAppenderators, name change of a method that was still using "legacy" rather than "openSegments"
Co-authored-by: Clint Wylie <cjwylie@gmail.com>
* Add error msg to parallel task's TaskStatus
* Consolidate failure block
* Add failure test
* Make it fail
* Add fail while stopped
* Simplify hash task test using a runner that fails after so many runs (parameter)
* Remove unthrown exception
* Use runner names to identify phase
* Added range partition kill test & fixed a timing bug with the custom runner
* Forbidden api
* Style
* Unit test code cleanup
* Added message to invalid state exception and improved readability of the phase error messages for the parallel task failure unit tests
This PR splits current SegmentLoader into SegmentLoader and SegmentCacheManager.
SegmentLoader - this class is responsible for building the segment object but does not expose any methods for downloading, cache space management, etc. Default implementation delegates the download operations to SegmentCacheManager and only contains the logic for building segments once downloaded. . This class will be used in SegmentManager to construct Segment objects.
SegmentCacheManager - this class manages the segment cache on the local disk. It fetches the segment files to the local disk, can clean up the cache, and in the future, support reserve and release on cache space. [See https://github.com/Make SegmentLoader extensible and customizable #11398]. This class will be used in ingestion tasks such as compaction, re-indexing where segment files need to be downloaded locally.
This PR refactors the code for QueryRunnerFactory#mergeRunners to accept a new interface called QueryProcessingPool instead of ExecutorService for concurrent execution of query runners. This interface will let custom extensions inject their own implementation for deciding which query-runner to prioritize first. The default implementation is the same as today that takes the priority of query into account. QueryProcessingPool can also be used as a regular executor service. It has a dedicated method for accepting query execution work so implementations can differentiate between regular async tasks and query execution tasks. This dedicated method also passes the QueryRunner object as part of the task information. This hook will let custom extensions carry any state from QuerySegmentWalker to QueryProcessingPool#mergeRunners which is not possible currently.
* Avoid mapping hydrants in create segments phase for native ingestion
* Drop queriable indices after a given sink is fully merged
* Do not drop memory mappings for realtime ingestion
* Style fixes
* Renamed to match use case better
* Rollback memoization code and use the real time flag instead
* Null ptr fix in FireHydrant toString plus adjustments to memory pressure tracking calculations
* Style
* Log some count stats
* Make sure sinks size is obtained at the right time
* BatchAppenderator unit test
* Fix comment typos
* Renamed methods to make them more readable
* Move persisted metadata from FireHydrant class to AppenderatorImpl. Removed superfluous differences and fix comment typo. Removed custom comparator
* Missing dependency
* Make persisted hydrant metadata map concurrent and better reflect the fact that keys are Java references. Maintain persisted metadata when dropping/closing segments.
* Replaced concurrent variables with normal ones
* Added batchMemoryMappedIndex "fallback" flag with default "false". Set this to "true" make code fallback to previous code path.
* Style fix.
* Added note to new setting in doc, using Iterables.size (and removing a dependency), and fixing a typo in a comment.
* Forgot to commit this edited documentation message
* Add ability to wait for segment availability for batch jobs
* IT updates
* fix queries in legacy hadoop IT
* Fix broken indexing integration tests
* address an lgtm flag
* spell checker still flagging for hadoop doc. adding under that file header too
* fix compaction IT
* Updates to wait for availability method
* improve unit testing for patch
* fix bad indentation
* refactor waitForSegmentAvailability
* Fixes based off of review comments
* cleanup to get compile after merging with master
* fix failing test after previous logic update
* add back code that must have gotten deleted during conflict resolution
* update some logging code
* fixes to get compilation working after merge with master
* reset interrupt flag in catch block after code review pointed it out
* small changes following self-review
* fixup some issues brought on by merge with master
* small changes after review
* cleanup a little bit after merge with master
* Fix potential resource leak in AbstractBatchIndexTask
* syntax fix
* Add a Compcation TuningConfig type
* add docs stipulating the lack of support by Compaction tasks for the new config
* Fixup compilation errors after merge with master
* Remove erreneous newline
* DruidInputSource: Fix issues in column projection, timestamp handling.
DruidInputSource, DruidSegmentReader changes:
1) Remove "dimensions" and "metrics". They are not necessary, because we
can compute which columns we need to read based on what is going to
be used by the timestamp, transform, dimensions, and metrics.
2) Start using ColumnsFilter (see below) to decide which columns we need
to read.
3) Actually respect the "timestampSpec". Previously, it was ignored, and
the timestamp of the returned InputRows was set to the `__time` column
of the input datasource.
(1) and (2) together fix a bug in which the DruidInputSource would not
properly read columns that are used as inputs to a transformSpec.
(3) fixes a bug where the timestampSpec would be ignored if you attempted
to set the column to something other than `__time`.
(1) and (3) are breaking changes.
Web console changes:
1) Remove "Dimensions" and "Metrics" from the Druid input source.
2) Set timestampSpec to `{"column": "__time", "format": "millis"}` for
compatibility with the new behavior.
Other changes:
1) Add ColumnsFilter, a new class that allows input readers to determine
which columns they need to read. Currently, it's only used by the
DruidInputSource, but it could be used by other columnar input sources
in the future.
2) Add a ColumnsFilter to InputRowSchema.
3) Remove the metric names from InputRowSchema (they were unused).
4) Add InputRowSchemas.fromDataSchema method that computes the proper
ColumnsFilter for given timestamp, dimensions, transform, and metrics.
5) Add "getRequiredColumns" method to TransformSpec to support the above.
* Various fixups.
* Uncomment incorrectly commented lines.
* Move TransformSpecTest to the proper module.
* Add druid.indexer.task.ignoreTimestampSpecForDruidInputSource setting.
* Fix.
* Fix build.
* Checkstyle.
* Misc fixes.
* Fix test.
* Move config.
* Fix imports.
* Fixup.
* Fix ShuffleResourceTest.
* Add import.
* Smarter exclusions.
* Fixes based on tests.
Also, add TIME_COLUMN constant in the web console.
* Adjustments for tests.
* Reorder test data.
* Update docs.
* Update docs to say Druid 0.22.0 instead of 0.21.0.
* Fix test.
* Fix ITAutoCompactionTest.
* Changes from review & from merging.
* druid task auto scale based on kafka lag
* fix kafkaSupervisorIOConfig and KinesisSupervisorIOConfig
* druid task auto scale based on kafka lag
* fix kafkaSupervisorIOConfig and KinesisSupervisorIOConfig
* test dynamic auto scale done
* auto scale tasks tested on prd cluster
* auto scale tasks tested on prd cluster
* modify code style to solve 29055.10 29055.9 29055.17 29055.18 29055.19 29055.20
* rename test fiel function
* change codes and add docs based on capistrant reviewed
* midify test docs
* modify docs
* modify docs
* modify docs
* merge from master
* Extract the autoScale logic out of SeekableStreamSupervisor to minimize putting more stuff inside there && Make autoscaling algorithm configurable and scalable.
* fix ci failed
* revert msic.xml
* add uts to test autoscaler create && scale out/in and kafka ingest with scale enable
* add more uts
* fix inner class check
* add IT for kafka ingestion with autoscaler
* add new IT in groups=kafka-index named testKafkaIndexDataWithWithAutoscaler
* review change
* code review
* remove unused imports
* fix NLP
* fix docs and UTs
* revert misc.xml
* use jackson to build autoScaleConfig with default values
* add uts
* use jackson to init AutoScalerConfig in IOConfig instead of Map<>
* autoscalerConfig interface and provide a defaultAutoScalerConfig
* modify uts
* modify docs
* fix checkstyle
* revert misc.xml
* modify uts
* reviewed code change
* reviewed code change
* code reviewed
* code review
* log changed
* do StringUtils.encodeForFormat when create allocationExec
* code review && limit taskCountMax to partitionNumbers
* modify docs
* code review
Co-authored-by: yuezhang <yuezhang@freewheel.tv>
* add offsetFetchPeriod to kinesis ingestion doc
* Remove jackson dependencies from extensions
* Use fixed delay for lag collection
* Metrics reset after finishing processing
* comments
* Broaden the list of exceptions to retry for
* Unit tests
* Add more tests
* Refactoring
* re-order metrics
* Doc suggestions
Co-authored-by: Charles Smith <38529548+techdocsmith@users.noreply.github.com>
* Add tests
Co-authored-by: Charles Smith <38529548+techdocsmith@users.noreply.github.com>
* Fix byte calculation for maxBytesInMemory to take into account of Sink/Hydrant Object overhead
* Fix byte calculation for maxBytesInMemory to take into account of Sink/Hydrant Object overhead
* Fix byte calculation for maxBytesInMemory to take into account of Sink/Hydrant Object overhead
* Fix byte calculation for maxBytesInMemory to take into account of Sink/Hydrant Object overhead
* fix checkstyle
* Fix byte calculation for maxBytesInMemory to take into account of Sink/Hydrant Object overhead
* Fix byte calculation for maxBytesInMemory to take into account of Sink/Hydrant Object overhead
* fix test
* fix test
* add log
* Fix byte calculation for maxBytesInMemory to take into account of Sink/Hydrant Object overhead
* address comments
* fix checkstyle
* fix checkstyle
* add config to skip overhead memory calculation
* add test for the skipBytesInMemoryOverheadCheck config
* add docs
* fix checkstyle
* fix checkstyle
* fix spelling
* address comments
* fix travis
* address comments
Today Kafka message support in streaming indexing tasks is limited to
message values, and does not provide a way to expose Kafka headers,
timestamps, or keys, which may be of interest to more specialized
Druid input formats. For instance, Kafka headers may be used to indicate
payload format/encoding or additional metadata, and timestamps are often
omitted from values in Kafka streams applications, since they are
included in the record.
This change proposes to introduce KafkaRecordEntity as InputEntity,
which would give input formats full access to the underlying Kafka record,
including headers, key, timestamps. It would also open access to low-level
information such as topic, partition, offset if needed.
KafkaEntity is a subclass of ByteEntity for backwards compatibility with
existing input formats, and to avoid introducing unnecessary complexity
for Kinesis indexing tasks.
* support multi-line text
* add test cases
* split json text into lines case by case
* improve exception handle
* fix CI
* use IntermediateRowParsingReader as base of JsonReader
* update doc
* ignore the non-immutable field in test case
* add more test cases
* mark `lineSplittable` as final
* fix testcases
* fix doc
* add a test case for SqlReader
* return all raw columns when exception occurs
* fix CI
* fix test cases
* resolve review comments
* handle ParseException returned by index.add
* apply Iterables.getOnlyElement
* fix CI
* fix test cases
* improve code in more graceful way
* fix test cases
* fix test cases
* add a test case to check multiple json string in one text block
* fix inspection check
* Introduce a Configurable Index Type
* Change to @UnstableApi
* Add AppendableIndexSpecTest
* Update doc
* Add spelling exception
* Add tests coverage
* Revert some of the changes to reduce diff
* Minor fixes
* Update getMaxBytesInMemoryOrDefault() comment
* Fix typo, remove redundant interface
* Remove off-heap spec (postponed to a later PR)
* Add javadocs to AppendableIndexSpec
* Describe testCreateTask()
* Add tests for AppendableIndexSpec within TuningConfig
* Modify hashCode() to conform with equals()
* Add comment where building incremental-index
* Add "EqualsVerifier" tests
* Revert some of the API back to AppenderatorConfig
* Don't use multi-line comments
* Remove knob documentation (deferred)
Gian Merlino
Kashif Faraz
AmatyaAvadhanula
Tejaswini Bandlamudi
Dr. Sizzles
Jonathan Wei
Clint Wylie
Adarsh Sanjeev
AmatyaAvadhanula
Jihoon Son
Parag Jain
Laksh Singla
Abhishek Agarwal
Agustin Gonzalez
zhangyue19921010
Yi Yuan
Lucas Capistrant
Maytas Monsereenusorn
Xavier Léauté
frank chen
Liran Funaro