* fix: restore support for unicode pipeline- and plugin-id's
JRuby's `Ruby#newSymbol(String)` throws an exception when provided a `String`
that contains characters outside of lower-ASCII because JRuby internals expect
"the incoming String to be one of our mangled ISO-8859-1 strings" as noted in
a comment on jruby/jruby#6217.
Instead, we use `Ruby#newString(String)` to create a new `RubyString` (which
works properly), and then rely on `RubyString#intern` to get our `RubySymbol`.
This fixes a regression introduced in the 8.7 series in which pipeline id's
are consistently represented as ruby symbols in the metrics store, and ensures
similar issue does not exist when specifying a plugin id that contains
characters above the lower-ASCII plane.
* fix: use properly-encoded RubySymbol in PipelineConfig
We cannot rely on `RubySymbol#toString` to produce a properly-encoded `String`
whe the string contains characters above the lower-ASCII plane because the
result is effectively a binary ruby-internal marshal of the bytes that only
holds when the symbol contains lower-ASCII.
Instead, we can use the internally-memoizing `RubySymbol#name` to get a
properly-encoded `RubyString`, and `RubyString#asJavaString()` to get a
properly-encoded java-`String`.
* fix: properly serialize unicode pipeline names in API output
Jackson's JSON serializer leaks the JRuby-internal byte structure of Symbols,
which only aligns with the byte-structure of the symbol's actual string when
that string is wholly-comprised of lower-ASCII characters.
By pre-converting Symbols to Strings, we ensure that the result is readable
and useful.
* spec: bypass monitoring specs for unicode pipeline ids when PQ enabled
(cherry picked from commit 0ec16ca398)
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
* Initial effort to initialize plugin flow metrics. Followings are addressed:
- Namespace store is shaped with RubySymbol key but filter and output codecs were using string key. This commit intends to standardize the namespace key with RubySymbol for filter & output codecs.
- Initializes throughput flow metrics for the input plugins.
- Initializes the worker cost per event and worker utilization for the filter and output plugins with only uptime metrics but it should combine with worker count, will be implemented in next commits.
- Fetching codec ID generated in ruby scope is possible but problematic to in Java scope. We will skip codec flow metrics since they are rarely produce the hard times.
* Worker utilization metrics implementation.
- Worker count will be provided as a fraction to the flow metrics. At the time when we fetch the metric value, fraction is applied.
* Unit tests added for fractured extended & simple metrics.
* Code review change requests applied.
- To simplify the scale (or fraction) at metric get value time, we can introduce the wrapper (`UpScaleMetric`) that applies the scale at metric value fetch time.
- Unit test added for `UpScaleMetric`
- We don't touch the codec namespace shape for now since we skipped codec metrics.
- Unused sources removed.
* Worker utilization and worker cost per event explanation added in the documentation.
* Integration test added for plugin-level flow metrics.
* Apply suggestions from code review
- Integration test failure fix: input plugin ID is not always in context config.
- Suggestions to simplify integration test source and rollback to intentional namings.
- Metrics explanation improvement in the doc.
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
* plugin flow: fix units; pass UptimeMetric and scale when needed
Aligns the units of the newly-introduced plugin metrics with the specification,
and passes our `UptimeMetric` through to the individual helper methods so that
they can scale appropriately for their context and our type-checker can ensure
we don't receive an incorrectly-scaled `Metric<Long>`.
Input `throughput`
------------------
all throughput metrics should be expressed in events-per-second; this
per-plugin scoped view of the pipeline's `input_throughput` flow should be
expressed in the same units.
Filters, Outputs `worker_utilization`
-------------------------------------
> a worker_utilization (duration / (uptime * worker count)) shows what percent
> of available resources an individual plugin instance is taking and can help
> identify where the blocker is.
To achieve this, we need to divide millis used by _millis_ available.
Filters, Outputs `worker_cost_per_event`
----------------------------------------
> we also provide a (to be named) cost-per-event metric (duration / event) to
> surface issues with a plugin that operates on a very small subset of events
> (via conditionals) but contributes disproportionately to the cost of getting
> its events through.
We start with a baseline of seconds-per-event, and acknowledge that this may
need to be scaled to a more understandable number before merging.
* plugin flow: express cost per event in millis per event
The "worker cost per event" metric when expressed as an inverse per-worker
throughput in seconds-per-event produces a range of values that are not
particularly easy to compare at-a-glance, with "nearly free" operations
being expressed in negative-exponent scientific notation and extremely
expensive operations being expressed with single-digits.
By scaling this metric up by a factor of 1000 to "millis per event" or its
eqivalent "seconds per thousand events", the resulting numbers in practice
are easier to make sense of:
+------------------------+--------------+---------------+------------+
| EXAMPLE / SCALE | s/event | ms/event | µs/event |
+------------------------+--------------+---------------+------------+
| no-op mutate @ 12k eps | 8.33e-05 | 0.0833 | 83.3 |
| stdout w/ dots codec | 0.000831 | 0.831 | 831 |
| ES out 1s RTT/125 | 0.008 | 8 | 8000 |
| ES out 30s retries/125 | 0.24 | 240 | 240000 |
| ES filter 1s/event | 1 | 1000 | 1000000 |
| grok 30s timeout | 30 | 30000 | 30000000 |
+------------------------+--------------+---------------+------------+
* plugin flow: reshape docs
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
Co-authored-by: Ry Biesemeyer <ry.biesemeyer@elastic.co>
* Collect growth events and bytes metrics if PQ is enabled: Java changes.
* Move queue flow under queue namespace.
* Pipeline level PQ flow metrics: add unit & integration tests.
* Include queue info in node stats sample.
* Apply suggestions from code review
Change uptime precision for PQ growth metrics to uptime seconds since PQ events are based on seconds.
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
* Add safeguard when using lazy delegating gauge type.
* flow metrics: simplify generics of lazy implementation
Enables interface `FlowMetrics::create` to take suppliers that _implement_
a `Metric<? extends Number>` instead of requiring them to be pre-cast, and
avoid unnecessary exposure of the metrics value-type into our lazy init.
* flow metrics: use lazy init for PQ gauge-based metrics
* noop: use enum equality
Avoids routing two enum values through `MetricType#toString()`
and `String#equals()` when they can be compared directly.
* Apply suggestions from code review
Optional.ofNullable used for safe return. Doc includes real tested expected metric values.
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
* flow metrics: make lazy-init wraper inherit from AbstractMetric
this allows the Jackson serialization annotations to work
* flow metrics: move pipeline queue-based flows into pipeline flow namespace
* Follow up for moving PQ growth metrics under pipeline.*.flow.
- Unit and integration tests are added or fixed.
- Documentation added along with sample response data
* flow: pipeline pq flow rates docs
* Do not expect flow in the queue section of API. Metrics moved to flow section.
Update logstash-core/spec/logstash/api/commands/stats_spec.rb
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
* Integration test failure fix.
Mistake: `flow_status` should be `pipeline_flow_stats`
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
* Integration test failures fix.
Number should be Numeric in the ruby specs.
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
* Make CI happy.
* api specs: use PQ only where needed
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
Co-authored-by: Ry Biesemeyer <ry.biesemeyer@elastic.co>
* Flow metrics: initial implementation (#14509)
* metrics: eliminate race condition when registering metrics
Ensure our fast-lookup and store tables cannot diverge in a race condition
by wrapping mutation of both in a single mutex and appropriately handle
another thread winning the race to the lock by using the value that it
persisted instead of writing our own.
* metrics: guard against intermediate namespace conflicts
- ensures our safeguard that prevents using an existing metric as a namespace
is applied to _intermediate_ nodes, not just the tail-node, eliminating a
potential crash when sending `fetch_or_store` to a metric object that is not
expected to respond to `fetch_or_store`.
- uses the atomic `Concurrent::Map#compute_if_absent` instead of the
non-atomic `Concurrent::Map#fetch_or_store`, which is prone to
last-write-wins during contention (as-written, this method is only
executed under lock and not subject to contention)
- uses `Enumerable#reduce` to eliminate the need for recursion
* flow: introduce auto-advancing UptimeMetric
* flow: introduce FlowMetric with minimal current/lifetime rates
* flow: initialize pipeline metrics at pipeline start
* Controller and service layer implementation for flow metrics. (#14514)
* Controller and service layer implementation for flow metrics.
* Add flow metrics to unit test and benchmark cli definitions.
* flow: fix tests for metric types to accomodate new one
* Renaming concurrency and backpressure metrics.
Rename `concurrency` to `worker_concurrency ` and `backpressure` to `queue_backpressure` to provide proper scope naming.
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
* metric: register flow metrics only when we have a collector (#14529)
the collector is absent when the pipeline is run in test with a
NullMetricExt, or when the pipeline is explicitly configured to
not collect metrics using `metric.collect: false`.
* Unit tests and integration tests added for flow metrics. (#14527)
* Unit tests and integration tests added for flow metrics.
* Node stat spec and pipeline spec metric updates.
* Metric keys statically imported, implicit error expectation added in metric spec.
* Fix node status API spec after renaming flow metrics.
* Removing flow metric from PipelinesInfo DS (used in peridoci metric snapshot), integration QA updates.
* metric: register flow metrics only when we have a collector (#14529)
the collector is absent when the pipeline is run in test with a
NullMetricExt, or when the pipeline is explicitly configured to
not collect metrics using `metric.collect: false`.
* Unit tests and integration tests added for flow metrics.
* Node stat spec and pipeline spec metric updates.
* Metric keys statically imported, implicit error expectation added in metric spec.
* Fix node status API spec after renaming flow metrics.
* Removing flow metric from PipelinesInfo DS (used in peridoci metric snapshot), integration QA updates.
* Rebasing with feature branch.
* metric: register flow metrics only when we have a collector
the collector is absent when the pipeline is run in test with a
NullMetricExt, or when the pipeline is explicitly configured to
not collect metrics using `metric.collect: false`.
* Apply suggestions from code review
Integration tests updated to test capturing the flow metrics.
* Flow metrics expectation updated in tegration tests.
* flow: refine integration expectations for reloads/monitoring
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
Co-authored-by: Ry Biesemeyer <ry.biesemeyer@elastic.co>
Co-authored-by: Mashhur <mashhur.sattorov@gmail.com>
* metric: add ScaledView with sub-unit precision to UptimeMetric (#14525)
* metric: add ScaledView with sub-unit precision to UptimeMetric
By presenting a _view_ of our metric that maintains sub-unit precision,
we prevent jitter that can be caused by our periodic poller not running at
exactly our configured cadence.
This is especially important as the UptimeMetric is used as the _denominator_ of
several flow metrics, and a capture at 4.999s that truncates to 4s, causes the
rate to be over-reported by ~25%.
The `UptimeMetric.ScaledView` implements `Metric<Number>`, so its full
lossless `BigDecimal` value is accessible to our `FlowMetric` at query time.
* metrics: reduce window for too-frequent-captures bug and document it
* fixup: provide mocked clock to flow metric
* Flow metrics cleanup (#14535)
* flow metrics: code-style and readability pass
* remove unused imports
* cleanup: simplify usage of internal helpers
* flow: migrate internals to use OptionalDouble
* Flow metrics global (#14539)
* flow: add global top-level flows
* docs: add flow metrics
* Top level flow metrics unit tests added. (#14540)
* Top level flow metrics unit tests added.
* Add unit tests when config reloads, make sure top-level flow metrics didn't get reset.
* Apply suggestions from code review
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
* Validating against Hash test cases updated.
* For the safety check against exact type in unit tests.
Co-authored-by: Ry Biesemeyer <yaauie@users.noreply.github.com>
* docs: section links and clarity in node stats API flow metrics
Co-authored-by: Mashhur <99575341+mashhurs@users.noreply.github.com>
Co-authored-by: Mashhur <mashhur.sattorov@gmail.com>
Fixes the issue #8752 in event.out counter. When a pipeline contains a drop filter the total out events counter should count only the events that reached the out stage.
This PR changes CompiledExecution.compute() interface to return the number of events that effectively reached the end of the pipeline. This change is used in WorkerLoop to update correctly the event.out metric, instead of relying on the batch's size.
Change DeadLetterQueueReader, so that if a missing segment file is
encountered at startup, the next valid entry will be used instead
Fixes#7433Fixes#7457
Add an initial set of metrics for the dead letter queue.
Metrics are supplied under the pipeline in the following format:
"dead_letter_queue": {
"queue_size_in_bytes": ...,
}
Metrics are populated via a PeriodicPoller
Also fixed up calculation of currentQueueSize to take account
of version headers, which was previously being skipped.
Additionally, whether the dlq is enabled, and if so, the path
of the dlq is supplied under the pipelines API endpoint
Resolves#7287Fixes#7338
The ability to change the log level as described at https://www.elastic.co/guide/en/logstash/current/logging.html#_logging_apis appears to no longer work. The root cause of this issue is that log4j2 context we initialize is different from the log4j2 context that we are setting via the API. The fix here is to mirror the functionality of org.apache.logging.log4j.core.config.Configurator.setLevel Java method, but instead use the logging_context initialized in our code via JRuby.
Fixes#7277Fixes#7321
* add multi_local source for multi pipelines
* introduce pipelines.yml
* introduce PipelineSettings class
* support reloading of pipeline parameters
* fix pipeline api call for _node/pipelines
* inform user pipelines.yml is ignored if -e or -f is enabled
in some tests against metrics api it's possible the values aren't there,
making the assertion fail. By default Stud.try only catches StandardError,
but rspec throws a RSpec::Expectations::ExpectationNotMetError on a failed
assertion, and this exception inherits directly from Exception.
This commit explicitly adds this exception to the list of Stud.try exceptions
Fixes#7177
because the metric subsystem may not be ready yet when the tests ask for
the values, the test code may generate a NoMethodError accessing nested
hashes.
The NoMethodError exception aborts the try mechanism instead of retrying.
This PR rescues potential exceptions and adds more expectations to
protect the try block.
Fixes#7074
Some integration spec tests fail to run
individually due missing require modules.
This fix it by ensuring all specs and helpers
require the needed modules.
Fixes#7037Fixes#7038
re #6508.
- removed `acked_count`, `unacked_count`, and migrated `unread_count` to
top-level `events` field.
- removed `current_size_in_bytes` info from queue node stats
Fixes#6510
