mirror of https://github.com/digital-asset/daml.git synced 2024-09-20 09:17:43 +03:00

Replace MetricsManager with MetricRegistry (#3574 )

Instead of passing around MetricsManager just for the timedFuture
method, we pass the underlying MetricRegistry directly and use
the timedFuture function as an actual function (passing in a timer metric).

The Sandbox still starts the JmxReporter, but the integrator of the Indexer
and the Ledger API Server needs to enable the reporting facilities themselves.

CHANGELOG_BEGIN

- [Ledger] Ledger API Server and Indexer now accept an instance of ``MetricRegistry`` as parameters. This gives implementors of ledger integrations the most flexibility to set up metrics reporting that works best for them.

CHANGELOG_END

2019-11-21 16:47:30 +01:00

4.4 KiB

Raw Blame History

ledger

Home of our reference ledger implementation (Sandbox) and various ledger related libraries.

Logging

Logging Configuration

The Sandbox and Ledger API Server use Logback for logging configuration.

Log Files

The Sandbox logs at INFO level to standard out and to the file sandbox.log in the current working directory.

Log levels

As most Java libraries and frameworks, the Sandbox and Ledger API Server use INFO as the default logging level. This level is for minimal and important information (usually only startup and normal shutdown events). INFO level logging should not produce increasing volume of logging during normal operation.

WARN level should be used for transition between healthy/unhealthy state, or in other close to error scenarios.

DEBUG level should be turned on only when investigating issues in the system, and usually that means we want the trail loggers. Normal loggers at DEBUG level can be useful sometimes (e.g. DAML interpretation).

Metrics

Sandbox and Ledger API Server provide a couple of useful metrics:

Sandbox and Ledger API Server

The Ledger API Server exposes basic metrics for all gRPC services and some additional ones.

Metric Name	Description
LedgerApi.com.digitalasset.ledger.api.v1.$SERVICE.$METHOD	A meter that tracks the number of calls to the respective service and method.
CommandSubmission.failedCommandInterpretations	A meter that tracks the failed command interpretations.
CommandSubmission.submittedTransactions	A timer that tracks the commands submitted to the backing ledger.

Indexer

Metric Name	Description
JdbcIndexer.processedStateUpdates	A timer that tracks duration of state update processing.
JdbcIndexer.lastReceivedRecordTime	A gauge that returns the last received record time in milliseconds since EPOCH.
JdbcIndexer.lastReceivedOffset	A gauge that returns that last received offset from the ledger.
JdbcIndexer.currentRecordTimeLag	A gauge that returns the difference between the Indexer's wallclock time and the last received record time in milliseconds.

Metrics Reporting

The Sandbox automatically makes all metrics available via JMX under the JMX domain com.digitalasset.platform.sandbox.

When building an Indexer or Ledger API Server the implementer/ledger integrator is responsible to set up a MetricRegistry and a suitable metric reporting strategy that fits their needs.

gRPC and back-pressure

RPC

Standard RPC requests should return with RESOURCE_EXHAUSTED status code to signal back-pressure. Envoy can be configured to retry on these errors. We have to be careful not to have any persistent changes when returning with such an error as the same original request can be retried on another service instance.

Streaming

gRPC's streaming protocol has built-in flow-control, but it's not fully active by default. What it does it controls the flow between the TCP/HTTP layer and the library so it builds on top of TCP's own flow control. The inbound flow control is active by default, but the outbound does not signal back-pressure out of the box.

AutoInboundFlowControl: The default behaviour for handling incoming items in a stream is to automatically signal demand after every onNext call. This is the correct thing to do if the handler logic is CPU bound and does not depend on other reactive downstream services. By default it's active on all inbound streams. One can disable this and signal demand by manually calling request to follow demands of downstream services. Disabling this feature is possible by calling disableAutoInboundFlowControl on CallStreamObserver.

ServerCallStreamObserver: casting an outbound StreamObserver manually to ServerCallStreamObserver gives us access to isReady and onReadyHandler. With these methods we can check if there is available capacity in the channel i.e. we are safe to push into it. This can be used to signal demand to our upstream flow. Note that gRPC buffers 32Kb data per channel and isReady will return false only when this buffer gets full.

4.4 KiB Raw Blame History