Streams Replication Manager monitoring and metrics
Streams Replication Manager metrics can give you insight into the status of SRM and the replication process as a whole. Learn how SRM metrics are collected, processed, and exposed on the SRM Service REST API.
Streams Replication Manager (SRM) consists of two main components (roles) that are each responsible for different tasks related to data replication. These are the SRM Driver and SRM Service roles.
The SRM Driver role is based on the Kafka Connect Framework and utilizes Connect connectors and connector tasks to replicate Kafka data. These connectors and tasks produce raw metrics that describe the replication process. The SRM service is responsible for collecting, processing, and exposing these metrics. The metrics can be queried using the SRM Service REST API or viewed in Streams Messaging Manager on the Replications page.
The following sections provide an overview on how raw SRM metrics are collected, how they are processed into metrics that can be used for monitoring.
Raw metric collection
Because the SRM Driver is built on top of the Kafka Connect framework, the connectors and connector tasks of the Driver produce raw Kafka Connect metrics. Although there are many metrics that exist within the Connect framework, SRM only collects the following raw metrics:
- Status metrics are collected from connectors.
- Status, byte-rate, replication-latency-ms, checkpoint-latency-ms, and replication-records-lag metrics are collected from connector tasks.
All of these metrics have two intrinsic properties. They have an origin and a subject. The origin denotes where the metric is collected from. The subject denotes what entity the metric describes. Based on these properties, the raw metrics can be organized into two groups. Connector-originated metrics, which describe the replication flows, and task-originated metrics, which describe the replication of individual topic partitions.
Metrics collected from connectors are connector-originated. Connectors within SRM’s framework represent the replications (also referred to as replication flows). Therefore, the subject of connector-originated metrics are the replications. Because the subject is the replication flow, connector-originated metrics have a wide scope and describe the replication process from a high level. Think of a two cluster replication scenario. The replication flow acts as a bridge between the two clusters. Connector-originated metrics describe the bridge, or in other words, the replication.
Metrics collected from connector tasks are task-originated. Tasks within SRM’s framework represent the actual replication work that is performed. The tasks are the processes that replicate the messages between the source and target clusters from the topic partitions. As a result, the subject of task-originated metrics is the topic partition. Compared to connector-originated metrics, task-originated metrics have a restricted scope and describe the replication process at a low level.
The raw metrics collected by SRM do not provide good insights on the replication process as a whole and cannot be used as is for monitoring. For example, even though connector-originated metrics describe the replications, a replication is always made up of multiple connector instances (or workers), which may even run on different nodes. Therefore, the raw metrics of a single connector instance (or worker) only provide partial data on the replication. Because of this, raw metrics are not exposed by the SRM service. Instead, SRM processes and aggregates them before exposing them for consumption on the REST API.
Metrics processing and aggregation
Evaluating a single metric is rarely useful. For example, looking at the transmitted bytes of an ongoing replication for a single topic partition in the last 30 seconds does not provide good insight on what is actually happening with the topic or the replication. To provide metrics that describe the replication process in a meaningful way, statistics derived from the raw data are required. In other words, the raw metric data must be processed before it can be exposed for consumption. Creating statistics involves gathering data, in this case raw metrics, and deriving their characteristics through processing and aggregation. The result of this are the metrics that SRM exposes on its REST API.
When processing raw data, SRM uses a combination of the following two strategies:
- SRM collects and processes metrics over a time period
- This is done because having access to the momentary status of a single entity does not provide real value. On the other hand, data gathered over time is descriptive and provides a good filter for temporary or short lived performance issues. SRM gathers data over time using sliding time-window averaging in the case of numeric metrics.
- SRM collects and processes metrics across different entities
- When monitoring SRM, the entities that you want to monitor are the topics and
replications. This is because it is the metrics of the topics and replications that can
give you tangible information on SRM’s status. However, there are no raw metrics
available that directly describe topics. Additionally, only status metrics are available
for the replications; raw metrics measuring byte rate, latency, or lag are not.
Nonetheless, SRM can still provide metrics on the topics and replications by way of
aggregating available raw metrics across different sets of entities.
Specifically, SRM gathers the raw metrics that describe the topic partitions (the task-originated raw metrics). Next, it aggregates the data over all topic partitions that belong to a topic. That is, over the topic. The resulting data are the metrics that describe the topic.
Metrics that describe the replications are produced in a similar way. SRM gathers the raw partition metrics and aggregates them over a certain replication. To be precise, over all topics that take part in that replication. The resulting data are the metrics that describe the replication.
All resulting metrics reported for the entities include the statistical minimum, maximum, average, and sum of the sample.
- Cluster-level metrics
- Cluster-level metrics are aggregated over a certain time frame and (contrary to the name) over a certain replication. These metrics reflect the overall performance of a certain replication.
- Topic-level metrics
- Topic-level metrics are aggregated over a certain time frame and over a topic. They reflect the overall progress and status of a replication with regards to a single topic.
Available numeric metrics
|byte-rate||cluster||The average number of bytes that are being replicated per second, aggregated over a replication|
|byte-rate||topic||The average number of bytes that are being replicated per second, aggregated over a topic|
|replication-latency-ms||cluster||Time it takes records to replicate from source to target cluster, aggregated over a replication|
|replication-latency-ms||topic||Time it takes records to replicate from source to target cluster, aggregated over a topic|
|checkpoint-latency-ms||cluster||Time it takes consumer group offsets to replicate from source to target cluster, aggregated over a replication|
|checkpoint-latency-ms||topic||Time it takes consumer group offsets to replicate from source to target cluster, aggregated over a topic|
|replication-records-lag||cluster||The number of “to be replicated” messages (difference between the source offset of the last replicated message and the log end offset of the source) aggregated over a replication. The “sum” statistics of the metric describes the overall lag of the replication with respect to a replication flow.|
|replication-records-lag||topic||The number of “to be replicated” messages (difference between the source offset of the last replicated message and the log end offset of the source) aggregated over a topic. The “sum” statistics of the metric describes the overall lag of the replication in a topic.|
Notes about calculation of metrics
When monitoring replication and checkpoint latency metrics, keep in mind the following regarding their calculation.
- Replication latency is the difference between the record timestamp of the message as saved in the source cluster and the time of the commit in the performing task. Because the source and target clusters in a replication are two distinct clusters, the record timestamp of the message and the commit time will come from two different clusters. As a result, the clocks of the two clusters must be in sync. Otherwise, SRM will calculate the latency incorrectly, which can lead to negative or other misleading values being reported
- Checkpoint latency is the difference between the time when the data necessary for creating a checkpoint is fetched (this includes consumer group offsets) and when the checkpoint is eventually committed.
Unlike raw metrics related to latency, record lag, or byte rate, raw status metrics are not numeric. These are simple strings of text that describe the state of the connector or task. For example, connectors will report that they are paused, running, or stopped. Connector tasks will report that they are running, stopped, or failed.
Similarly to other raw metrics, status metrics are not exposed directly by SRM and are also processed and aggregated. SRM uses status metrics as input to create detailed status messages that describe the overall status of a replication and as input for health checks.
The status messages that SRM produces are more complex than the raw metrics used for input. They include the reason and cause of why a specific replication is in the reported state. These messages can be viewed on the Replications page of the SMM UI.