Sysmonitor stats (mem, fs, disk, net, cpu, swap, sys, tcp) are reported by all Druid processes, including Peons that are ephemeral in nature. Since Peons always run on the same host as the MiddleManager that spawned them and is unlikely to change, the SyMonitor metrics emitted by Peon are merely duplicates. This is often not a problem except when machines are super-beefy. Imagine a 64-core machine and 32 workers running on this machine. now you will have each Peon reporting metrics for each core. that's an increase of (32 * 64)x in the number of metrics. This leads to a metric explosion.
This PR updates MetricsModule to check node role running while registering SysMonitor and not to load any existing SysMonitor$Stats.
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.
* fix bug in ObjectFlatteners.toMap which caused null values in avro-stream/avro-ocf/parquet/orc to be converted to {} instead of null
* fix parquet test that expected wrong behavior, my bad heh
* Frame format for data transfer and short-term storage.
As we move towards query execution plans that involve more transfer
of data between servers, it's important to have a data format that
provides for doing this more efficiently than the options available to
us today.
This patch adds:
- Columnar frames, which support fast querying.
- Row-based frames, which support fast sorting via memory comparison
and fast whole-row copies via memory copying.
- Frame files, a container format that can be stored on disk or
transferred between servers.
The idea is we should use row-based frames when data is expected to
be sorted, and columnar frames when data is expected to be queried.
The code in this patch is not used in production yet. Therefore, the
patch involves minimal changes outside of the org.apache.druid.frame
package. The main ones are adjustments to SqlBenchmark to add benchmarks
for queries on frames, and the addition of a "forEach" method to Sequence.
* Fixes based on tests, static analysis.
* Additional fixes.
* Skip DS mapping tests on JDK 14+
* Better JDK checking in tests.
* Fix imports.
* Additional comment.
* Adjustments from code review.
* Update test case.
* jvm gc to mxbeans
* add zgc and shenandoah #12476
* remove tryCreateGcCounter
* separate the space collector
* blend GcGenerationCollector into GcCollector
* add jdk surefire argLine
* 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
In a heterogeneous environment, sometimes you don't have control over the input folder. Upstream can put any folder they want. In this situation the S3InputSource.java is unusable.
Most people like me solved it by using Airflow to fetch the full list of parquet files and pass it over to Druid. But doing this explodes the JSON spec. We had a situation where 1 of the JSON spec is 16MB and that's simply too much for Overlord.
This patch allows users to pass {"filter": "*.parquet"} and let Druid performs the filtering of the input files.
I am using the glob notation to be consistent with the LocalFirehose syntax.
This commit contains the cleanup needed for the new integration test framework.
Changes:
- Fix log lines, misspellings, docs, etc.
- Allow the use of some of Druid's "JSON config" objects in tests
- Fix minor bug in `BaseNodeRoleWatcher`
* 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.
The "exceptionCaught" handler may get called multiple times. We should
only return the channel to the pool the first time. Returning it more
than once leads to a warning like "Resource at key[%s] was returned
multiple times?"
* Direct UTF-8 access for "in" filters.
Directly related:
1) InDimFilter: Store stored Strings (in ValuesSet) plus sorted UTF-8
ByteBuffers (in valuesUtf8). Use valuesUtf8 whenever possible. If
necessary, the input set is copied into a ValuesSet. Much logic is
simplified, because we always know what type the values set will be.
I think that there won't even be an efficiency loss in most cases.
InDimFilter is most frequently created by deserialization, and this
patch updates the JsonCreator constructor to deserialize
directly into a ValuesSet.
2) Add Utf8ValueSetIndex, which InDimFilter uses to avoid UTF-8 decodes
during index lookups.
3) Add unsigned comparator to ByteBufferUtils and use it in
GenericIndexed.BYTE_BUFFER_STRATEGY. This is important because UTF-8
bytes can be compared as bytes if, and only if, the comparison
is unsigned.
4) Add specialization to GenericIndexed.singleThreaded().indexOf that
avoids needless ByteBuffer allocations.
5) Clarify that objects returned by ColumnIndexSupplier.as are not
thread-safe. DictionaryEncodedStringIndexSupplier now calls
singleThreaded() on all relevant GenericIndexed objects, saving
a ByteBuffer allocation per access.
Also:
1) Fix performance regression in LikeFilter: since #12315, it applied
the suffix matcher to all values in range even for type MATCH_ALL.
2) Add ObjectStrategy.canCompare() method. This fixes LikeFilterBenchmark,
which was broken due to calls to strategy.compare in
GenericIndexed.fromIterable.
* Add like-filter implementation tests.
* Add in-filter implementation tests.
* Add tests, fix issues.
* Fix style.
* Adjustments from review.
* Ensure ByteBuffers allocated in tests get freed.
Many tests had problems where a direct ByteBuffer would be allocated
and then not freed. This is bad because it causes flaky tests.
To fix this:
1) Add ByteBufferUtils.allocateDirect(size), which returns a ResourceHolder.
This makes it easy to free the direct buffer. Currently, it's only used
in tests, because production code seems OK.
2) Update all usages of ByteBuffer.allocateDirect (off-heap) in tests either
to ByteBuffer.allocate (on-heap, which are garbaged collected), or to
ByteBufferUtils.allocateDirect (wherever it seemed like there was a good
reason for the buffer to be off-heap). Make sure to close all direct
holders when done.
* Changes based on CI results.
* A different approach.
* Roll back BitmapOperationTest stuff.
* Try additional surefire memory.
* Revert "Roll back BitmapOperationTest stuff."
This reverts commit 49f846d9e3.
* Add TestBufferPool.
* Revert Xmx change in tests.
* Better behaved NestedQueryPushDownTest. Exit tests on OOME.
* Fix TestBufferPool.
* Remove T1C from ARM tests.
* Somewhat safer.
* Fix tests.
* Fix style stuff.
* Additional debugging.
* Reset null / expr configs better.
* ExpressionLambdaAggregatorFactory thread-safety.
* Alter forkNode to try to get better info when a JVM crashes.
* Fix buffer retention in ExpressionLambdaAggregatorFactory.
* Remove unused import.
Allow a Druid cluster to kill segments whose interval_end is a date in the future. This can be done by setting druid.coordinator.kill.durationToRetain to a negative period. For example PT-24H would allow segments to be killed if their interval_end date was 24 hours or less into the future at the time that the kill task is generated by the system.
A cluster operator can also disregard the druid.coordinator.kill.durationToRetain entirely by setting a new configuration, druid.coordinator.kill.ignoreDurationToRetain=true. This ignores interval_end date when looking for segments to kill, and instead is capable of killing any segment marked unused. This new configuration is off by default, and a cluster operator should fully understand and accept the risks if they enable it.
* GroupBy: Reduce allocations by reusing entry and key holders.
Two main changes:
1) Reuse Entry objects returned by various implementations of
Grouper.iterator.
2) Reuse key objects contained within those Entry objects.
This is allowed by the contract, which states that entries must be
processed and immediately discarded. However, not all call sites
respected this, so this patch also updates those call sites.
One particularly sneaky way that the old code retained entries too long
is due to Guava's MergingIterator and CombiningIterator. Internally,
these both advance to the next value prior to returning the current
value. So, this patch addresses that in two ways:
1) For merging, we have our own implementation MergeIterator already,
although it had the same problem. So, this patch updates our
implementation to return the current item prior to advancing to the
next item. It also adds a forbidden-api entry to ensure that this
safer implementation is used instead of Guava's.
2) For combining, we address the problem in a different way: by copying
the key when creating the new, combined entry.
* Attempt to fix test.
* Remove unused import.
* DimensionRangeShardSpec speed boost.
Calling isEmpty() and equals() on RangeSets is expensive, because these
fall back on default implementations that call size(). And size() is
_also_ a default implementation that iterates the entire collection.
* Fix and test from code review.
* Reduce allocations due to Jackson serialization.
This patch attacks two sources of allocations during Jackson
serialization:
1) ObjectMapper.writeValue and JsonGenerator.writeObject create a new
DefaultSerializerProvider instance for each call. It has lots of
fields and creates pressure on the garbage collector. So, this patch
adds helper functions in JacksonUtils that enable reuse of
SerializerProvider objects and updates various call sites to make
use of this.
2) GroupByQueryToolChest copies the ObjectMapper for every query to
install a special module that supports backwards compatibility with
map-based rows. This isn't needed if resultAsArray is set and
all servers are running Druid 0.16.0 or later. This release was a
while ago. So, this patch disables backwards compatibility by default,
which eliminates the need to copy the heavyweight ObjectMapper. The
patch also introduces a configuration option that allows admins to
explicitly enable backwards compatibility.
* Add test.
* Update additional call sites and add to forbidden APIs.
* JvmMonitor: Handle more generation and collector scenarios.
ZGC on Java 11 only has a generation 1 (there is no 0). This causes
a NullPointerException when trying to extract the spacesCount for
generation 0. In addition, ZGC on Java 15 has a collector number 2
but no spaces in generation 2, which breaks the assumption that
collectors always have same-numbered spaces.
This patch adjusts things to be more robust, enabling the JvmMonitor
to work properly for ZGC on both Java 11 and 15.
* Test adjustments.
* Improve surefire arglines.
* Need a placeholder
This PR enables ARM builds on Travis. I've ported over the changes from @martin-g on reducing heap requirements for some of the tests to ensure they run well on Travis arm instances.
GCP allows bucket names to contain underscores. When a location in such a bucket
is mapped to `java.net.URI`, `URI.getHost()` returns null. `URI.getHost()` is used as
the bucket name in `CloudObjectLocation`, leading to an NPE.
This commit uses `URI.getAuthority()` as the bucket name if `URI.getHost()` is null.
* 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
If there are many shards, mapper of IndexGeneratorJob seems to spend a lot of time in calling
DimensionRangeShardSpec.isInChunk to lookup target shard. This can be significantly improved
by using binary search instead of comparing an input row to every shardSpec.
Changes:
* Add `BaseDimensionRangeShardSpec` which provides a binary-search-based
implementation for `createLookup`
* `DimensionRangeShardSpec`, `SingleDimensionShardSpec`, and
`DimensionRangeBucketShardSpec` now extend `BaseDimensionRangeShardSpec`
* 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
* Counting nulls in String cardinality with a config
* Adding tests for the new config
* Wrapping the vectorize part to allow backward compatibility
* Adding different tests, cleaning the code and putting the check at the proper position, handling hasRow() and hasValue() changes
* Updating testcase and code
* Adding null handling test to improve coverage
* Checkstyle fix
* Adding 1 more change in docs
* Making docs clearer
* remove use of reflection in EnvironmentVariableDynamicConfigProvider for Java 17 compatibility
* fix mocks mock objects not getting closed properly, causing issues with Java 17
* remove use of deprecated methods and rules in tests
* finds complete and active tasks from the same snapshot
* overlord resource
* unit test
* integration test
* javadoc and cleanup
* more cleanup
* fix test and add more
The latest version of Error Prone now requires Java 11. Upgrading means we can
remove a lot of the maven profile complexity required to run checks with Java 8.
This also requires switching our strict build to use Java 11.
* update error-prone to 2.11
* remove need for specific maven profiles for Java 8 and Java 15
* fix additional Error Prone warnings with Java 11
* update strict build to use Java 11
Add config for eager / lazy connection initialization in ResourcePool
Description
Currently, when multiple tasks are launched, each of them eagerly initializes a full pool's worth of connections to the coordinator.
While this is acceptable when the parameter for number of eagerConnections (== maxSize) is small, this can be problematic in environments where it's a large value (say 1000) and multiple tasks are launched simultaneously, which can cause a large number of connections to be created to the coordinator, thereby overwhelming it.
Patch
Nodes like the broker may require eager initialization of resources and do not create connections with the Coordinator.
It is unnecessary to do this with other types of nodes.
A config parameter eagerInitialization is added, which when set to true, initializes the max permissible connections when ResourcePool is initialized.
If set to false, lazy initialization of connection resources takes place.
NOTE: All nodes except the broker have this new parameter set to false in the quickstart as part of this PR
Algorithm
The current implementation relies on the creation of maxSize resources eagerly.
The new implementation's behaviour is as follows:
If a resource has been previously created and is available, lend it.
Else if the number of created resources is less than the allowed parameter, create and lend it.
Else, wait for one of the lent resources to be returned.
* Tombstone support for replace functionality
* A used segment interval is the interval of a current used segment that overlaps any of the input intervals for the spec
* Update compaction test to match replace behavior
* Adapt ITAutoCompactionTest to work with tombstones rather than dropping segments. Add support for tombstones in the broker.
* Style plus simple queriableindex test
* Add segment cache loader tombstone test
* Add more tests
* Add a method to the LogicalSegment to test whether it has any data
* Test filter with some empty logical segments
* Refactor more compaction/dropexisting tests
* Code coverage
* Support for all empty segments
* Skip tombstones when looking-up broker's timeline. Discard changes made to tool chest to avoid empty segments since they will no longer have empty segments after lookup because we are skipping over them.
* Fix null ptr when segment does not have a queriable index
* Add support for empty replace interval (all input data has been filtered out)
* Fixed coverage & style
* Find tombstone versions from lock versions
* Test failures & style
* Interner was making this fail since the two segments were consider equal due to their id's being equal
* Cleanup tombstone version code
* Force timeChunkLock whenever replace (i.e. dropExisting=true) is being used
* Reject replace spec when input intervals are empty
* Documentation
* Style and unit test
* Restore test code deleted by mistake
* Allocate forces TIME_CHUNK locking and uses lock versions. TombstoneShardSpec added.
* Unused imports. Dead code. Test coverage.
* Coverage.
* Prevent killer from throwing an exception for tombstones. This is the killer used in the peon for killing segments.
* Fix OmniKiller + more test coverage.
* Tombstones are now marked using a shard spec
* Drop a segment factory.json in the segment cache for tombstones
* Style
* Style + coverage
* style
* Add TombstoneLoadSpec.class to mapper in test
* Update core/src/main/java/org/apache/druid/segment/loading/TombstoneLoadSpec.java
Typo
Co-authored-by: Jonathan Wei <jon-wei@users.noreply.github.com>
* Update docs/configuration/index.md
Missing
Co-authored-by: Jonathan Wei <jon-wei@users.noreply.github.com>
* Typo
* Integrated replace with an existing test since the replace part was redundant and more importantly, the test file was very close or exceeding the 10 min default "no output" CI Travis threshold.
* Range does not work with multi-dim
Co-authored-by: Jonathan Wei <jon-wei@users.noreply.github.com>
* Always reopen stream in FileUtils.copyLarge, RetryingInputStream.
When an InputStream throws an exception from one of its read methods,
we should assume it's bad and reopen it.
The main changes here are:
- In FileUtils.copyLarge, replace InputStream with InputStreamSupplier.
- In RetryingInputStream, collapse retryCondition and resetCondition
into a single condition. Also, make it required, since every usage
is passing in a specific condition anyway.
* Test fixes.
* Fix read impl.