This commit fixes issues with delayed supervisor termination during certain transient states.
Tasks can be created during supervisor termination and left behind since the cleanup may
not consider these newly added tasks.
#12178 added a lock for the entire process of task creation to prevent such dangling tasks.
But it also introduced a deadlock scenario as follows:
- An invocation of `runInternal` is in progress.
- A `stop` request comes, acquires `stateChangeLock` and submit a `ShutdownNotice`
- `runInternal` keeps waiting to acquire the `stateChangeLock`
- `ShutdownNotice` remains stuck in the notice queue because `runInternal` is still running
- After some timeout, the supervisor goes through a forced termination
Fix:
* `SeekableStreamSupervisor.runInternal` - do not try to acquire lock if supervisor is already stopping
* `SupervisorStateManager.maybeSetState` - do not allow transitions from STOPPING state
* Fixing RACE in HTTP remote task Runner
* Changes in the interface
* Updating documentation
* Adding test cases to SwitchingTaskLogStreamer
* Adding more tests
* Fix serialization in TaskReportFileWriters.
For some reason, serializing a Map<String, TaskReport> would omit the
"type" field. Explicitly sending each value through the ObjectMapper
fixes this, because the type information does not get lost.
* Fixes for static analysis.
* Fix race in TaskQueue.notifyStatus.
It was possible for manageInternal to relaunch a task while it was
being cleaned up, due to a race that happens when notifyStatus is
called to clean up a successful task:
1) In a critical section, notifyStatus removes the task from "tasks".
2) Outside a critical section, notifyStatus calls taskRunner.shutdown
to let the task runner know it can clear out its data structures.
3) In a critical section, syncFromStorage adds the task back to "tasks",
because it is still present in metadata storage.
4) In a critical section, manageInternalCritical notices that the task
is in "tasks" and is not running in the taskRunner, so it launches
it again.
5) In a (different) critical section, notifyStatus updates the metadata
store to set the task status to SUCCESS.
6) The task continues running even though it should not be.
The possibility for this race was introduced in #12099, which shrunk
the critical section in notifyStatus. Prior to that patch, a single
critical section encompassed (1), (2), and (5), so the ordering above
was not possible.
This patch does the following:
1) Fixes the race by adding a recentlyCompletedTasks set that prevents
the main management loop from doing anything with tasks that are
currently being cleaned up.
2) Switches the order of the critical sections in notifyStatus, so
metadata store updates happen first. This is useful in case of
server failures: it ensures that if the Overlord fails in the midst
of notifyStatus, then completed-task statuses are still available in
ZK or on MMs for the next Overlord. (Those are cleaned up by
taskRunner.shutdown, which formerly ran first.) This isn't related
to the race described above, but is fixed opportunistically as part
of the same patch.
3) Changes the "tasks" list to a map. Many operations require retrieval
or removal of individual tasks; those are now O(1) instead of O(N)
in the number of running tasks.
4) Changes various log messages to use task ID instead of full task
payload, to make the logs more readable.
* Fix format string.
* Update comment.
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.
* Refactor Guice initialization
Builders for various module collections
Revise the extensions loader
Injector builders for server startup
Move Hadoop init to indexer
Clean up server node role filtering
Calcite test injector builder
* Revisions from review comments
* Build fixes
* Revisions from review comments
Few indexing tasks register RealtimeMetricsMonitor or TaskRealtimeMetricsMonitor with the process’s MonitorScheduler when they start. These monitors never unregister themselves (they always return true, they'd need to return false to unregister). Each of these monitors emits a set of metrics once every druid.monitoring.emissionPeriod.
As a result, after executing several tasks for a while, Indexer emits metrics of these tasks even after they're long gone.
Proposed Solution
Since one should be able to obtain the last round of ingestion metrics after the task unregisters the monitor, introducing lastRoundMetricsToBePushed variable to keep track of the same and overriding the AbstractMonitor.monitor method in RealtimeMetricsMonitor, TaskRealtimeMetricsMonitor to implement the new logic.
* Add EIGHT_HOUR into possible list of Granularities.
* Add the missing definition.
* fix test.
* Fix another test.
* Stylecheck finally passed.
Co-authored-by: Didip Kerabat <didip@apple.com>
* 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.
* Poison StupidPool and fix resource leaks
There are various resource leaks from test setup as well as some
corners in query processing. We poison the StupidPool to start failing
tests when the leaks come and fix any issues uncovered from that so
that we can start from a clean baseline.
Unfortunately, because of how poisoning works,
we can only fail future checkouts from the same pool,
which means that there is a natural race between a
leak happening -> GC occurs -> leak detected -> pool poisoned.
This race means that, depending on interleaving of tests,
if the very last time that an object is checked out
from the pool leaks, then it won't get caught.
At some point in the future, something will catch it,
however and from that point on it will be deterministic.
* Remove various things left over from iterations
* Clean up FilterAnalysis and add javadoc on StupidPool
* Revert changes to .idea/misc.xml that accidentally got pushed
* Style and test branches
* Stylistic woes
The expiry timeout is compared against the current time but the condition is reversed.
This means that as soon as a supervisor task finishes, its partitions are cleaned up,
irrespective of the specified `intermediaryPartitionTimeout` period.
After these changes, the `intermediaryPartitionTimeout` will start getting honored.
Changes
* Fix the condition
* Add tests to verify the new correct behaviour
* Reduce the default expiry timeout from P1D to PT5M
to retain current behaviour in case of default configs.
* Fix flaky KafkaIndexTaskTest.
The testRunTransactionModeRollback case had many race conditions. Most notably,
it would commit a transaction and then immediately check to see that the results
were *not* indexed. This is racey because it relied on the indexing thread being
slower than the test thread.
Now, the case waits for the transaction to be processed by the indexing thread
before checking the results.
* Changes from review.
* Remove null and empty fields from native queries
* Test fixes
* Attempted IT fix.
* Revisions from review comments
* Build fixes resulting from changes suggested by reviews
* IT fix for changed segment size
The web-console (indirectly) calls the Overlord’s GET tasks API to fetch the tasks' summary which in turn queries the metadata tasks table. This query tries to fetch several columns, including payload, of all the rows at once. This introduces a significant memory overhead and can cause unresponsiveness or overlord failure when the ingestion tab is opened multiple times (due to several parallel calls to this API)
Another thing to note is that the task table (the payload column in particular) can be very large. Extracting large payloads from such tables can be very slow, leading to slow UI. While we are fixing the memory pressure in the overlord, we can also fix the slowness in UI caused by fetching large payloads from the table. Fetching large payloads also puts pressure on the metadata store as reported in the community (Metadata store query performance degrades as the tasks in druid_tasks table grows · Issue #12318 · apache/druid )
The task summaries returned as a response for the API are several times smaller and can fit comfortably in memory. So, there is an opportunity here to fix the memory usage, slow ingestion, and under-pressure metadata store by removing the need to handle large payloads in every layer we can. Of course, the solution becomes complex as we try to fix more layers. With that in mind, this page captures two approaches. They vary in complexity and also in the degree to which they fix the aforementioned problems.
* ForkingTaskRunner: Set ActiveProcessorCount for tasks.
This prevents various automatically-sized thread pools from being unreasonably
large (we don't want each task to size its pools as if it is the only thing on
the entire machine).
* Fix tests.
* Add missing LifecycleStart annotation.
* ForkingTaskRunner needs ManageLifecycle.
Often users are submitting queries, and ingestion specs that work only if the relevant extension is not loaded. However, the error is too technical for the users and doesn't suggest them to check for missing extensions. This PR modifies the error message so users can at least check their settings before assuming that the error is because of a bug.
* Emit state of replace and append for native batch tasks
* Emit count of one depending on batch ingestion mode (APPEND, OVERWRITE, REPLACE)
* Add metric to compaction job
* Avoid null ptr exc when null emitter
* Coverage
* Emit tombstone & segment counts
* Tasks need a type
* Spelling
* Integrate BatchIngestionMode in batch ingestion tasks functionality
* Typos
* Remove batch ingestion type from metric since it is already in a dimension. Move IngestionMode to AbstractTask to facilitate having mode as a dimension. Add metrics to streaming. Add missing coverage.
* Avoid inner class referenced by sub-class inspection. Refactor computation of IngestionMode to make it more robust to null IOConfig and fix test.
* Spelling
* Avoid polluting the Task interface
* Rename computeCompaction methods to avoid ambiguous java compiler error if they are passed null. Other minor cleanup.
* Deal with potential cardinality estimate being negative and add logging
* Fix typo in name
* Refine and minimize logging
* Make it info based on code review
* Create a named constant for the magic number
Issue:
Even though `CompactionTuningConfig` allows a `maxColumnsToMerge` config
(to optimize memory usage, particulary for datasources with many dimensions),
the corresponding client object `ClientCompactionTaskQueryTuningConfig`
(used by the coordinator duty `CompactSegments` to trigger auto-compaction)
does not contain this field. Thus, the value of `maxColumnsToMerge` specified
in any datasource compaction config is ignored.
Changes:
- Add field `maxColumnsToMerge` in `ClientCompactionTaskQueryTuningConfig`
and `UserCompactionTaskQueryTuningConfig`
- Fix tests
* RemoteTaskRunner: Fix NPE in streamTaskReports.
It is possible for a work item to drop out of runningTasks after the
ZkWorker is retrieved. In this case, the current code would throw
an NPE.
* Additional tests and additional fixes.
* Fix import.
1) Align "Assigning task" log messages between RTR and HRTR.
2) Remove confusing reference to "Coordinator".
3) Move "Not assigning task" message from INFO to DEBUG. It's not super
important to see this message: we mainly want to see what _does_ get
assigned.
4) Reword "Task switched from pending to running" message to better
match the structure of the "Assigning task" message from the same
method.
* Add builder for TaskToolbox.
The main purpose of this change is to make it easier to create
TaskToolboxes in tests. However, the builder is used in production
too, by TaskToolboxFactory.
* Fix imports, adjust formatting.
* Fix import.
Currently all Druid processes share the same log4j2 configuration file located in _common directory. Since peon processes are spawned by middle manager process, they derivate the environment variables from the middle manager. These variables include those in the log4j2.xml controlling to which file the logger writes the log.
But current task logging mechanism requires the peon processes to output the log to console so that the middle manager can redirect the console output to a file and upload this file to task log storage.
So, this PR imposes this requirement to peon processes, whatever the configuration is in the shared log4j2.xml, peon processes always write the log to console.
* concurrency: introduce GuardedBy to TaskQueue
* perf: Introduce TaskQueueScaleTest to test performance of TaskQueue with large task counts
This introduces a test case to confirm how long it will take to launch and manage (aka shutdown)
a large number of threads in the TaskQueue.
h/t to @gianm for main implementation.
* perf: improve scalability of TaskQueue with large task counts
* linter fixes, expand test coverage
* pr feedback suggestion; swap to different linter
* swap to use SuppressWarnings
* Fix TaskQueueScaleTest.
Co-authored-by: Gian Merlino <gian@imply.io>
Following up on #12315, which pushed most of the logic of building ImmutableBitmap into BitmapIndex in order to hide the details of how column indexes are implemented from the Filter implementations, this PR totally refashions how Filter consume indexes. The end result, while a rather dramatic reshuffling of the existing code, should be extraordinarily flexible, eventually allowing us to model any type of index we can imagine, and providing the machinery to build the filters that use them, while also allowing for other column implementations to implement the built-in index types to provide adapters to make use indexing in the current set filters that Druid provides.
This PR is to measure how long a task stays in the pending queue and emits the value with the metric task/pending/time. The metric is measured in RemoteTaskRunner and HttpRemoteTaskRunner.
An example of the metric:
```
2022-04-26T21:59:09,488 INFO [rtr-pending-tasks-runner-0] org.apache.druid.java.util.emitter.core.LoggingEmitter - {"feed":"metrics","timestamp":"2022-04-26T21:59:09.487Z","service":"druid/coordinator","host":"localhost:8081","version":"2022.02.0-iap-SNAPSHOT","metric":"task/pending/time","value":8,"dataSource":"wikipedia","taskId":"index_parallel_wikipedia_gecpcglg_2022-04-26T21:59:09.432Z","taskType":"index_parallel"}
```
------------------------------------------
Key changed/added classes in this PR
Emit metric task/pending/time in classes RemoteTaskRunner and HttpRemoteTaskRunner.
Update related factory classes and tests.
* Make tombstones ingestible by having them return an empty result set.
* Spotbug
* Coverage
* Coverage
* Remove unnecessary exception (checkstyle)
* Fix integration test and add one more to test dropExisting set to false over tombstones
* Force dropExisting to true in auto-compaction when the interval contains only tombstones
* Checkstyle, fix unit test
* Changed flag by mistake, fixing it
* Remove method from interface since this method is specific to only DruidSegmentInputentity
* Fix typo
* Adapt to latest code
* Update comments when only tombstones to compact
* Move empty iterator to a new DruidTombstoneSegmentReader
* Code review feedback
* Checkstyle
* Review feedback
* Coverage
* Optionally load segment index files into page cache on bootstrap and new segment download
* Fix unit test failure
* Fix test case
* fix spelling
* fix spelling
* fix test and test coverage issues
Co-authored-by: Jian Wang <wjhypo@gmail.com>
* add impl
* add impl
* fix checkstyle
* add impl
* add unit test
* fix stuff
* fix stuff
* fix stuff
* add unit test
* add more unit tests
* add more unit tests
* add IT
* add IT
* add IT
* add IT
* add ITs
* address comments
* fix test
* fix test
* fix test
* address comments
* address comments
* address comments
* fix conflict
* fix checkstyle
* address comments
* fix test
* fix checkstyle
* fix test
* fix test
* fix IT
The `javaOpts` property is being read from task context but not `javaOptsArray`.
Changes:
- Read `javaOptsArray` from task context in `ForkingTaskRunner`.
- Add test to verify that `javaOptsArray` in task context takes precedence over `javaOpts`
* 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
Parallel indexing with range partitioning can often cause OOM in the
`ParallelIndexSupervisorTask` during the dimension distribution phase.
This typically happens because of too many `StringSketch` objects
obtained from the different `partial_dimension_distribution` sub-tasks.
We need not keep any of the sketches in memory until we need to compute
the PartitionBoundaries for the respective interval.
Changes
- Extract `StringDistribution` from `DimensionDistributionReport`s when they are received
and write to disk inside the task/temp/distributions
- After all the subtasks have finished, iterate over all the intervals one by one
- For each interval, read the distributions from disk, merge them and create `PartitionBoundaries`.
- Cleanup task/temp/distributions directory when all `PartitionBoundaries` have been determined
Gian Merlino
AmatyaAvadhanula
Abhishek Agarwal
Karan Kumar
Clint Wylie
Adarsh Sanjeev
Herb Brewer
Paul Rogers
刘小辉
Maytas Monsereenusorn
Tejaswini Bandlamudi
Rohan Garg
Didip Kerabat
imply-cheddar
Kashif Faraz
Agustin Gonzalez
superivaj
Frank Chen
Jason Koch
zachjsh
Rocky Chen
Jihoon Son
Parag Jain
AmatyaAvadhanula