Alert metrics

Every figure an alert rule can watch, by platform: the name the rule dialog shows, the key it is stored under, its unit and what it lands on. A rule's metric is chosen from this list and nothing else — a key with no figure behind it would be a rule that saves and never fires.

What a metric lands on decides what a rule's scope expands to. A metric that lands on a kind of resource — a topic, a consumer group, a queue, a connector — measures each one in the scope, or the one a resource-scoped rule names. A metric that lands on the connection has one figure per cluster. One marked per node is measured on each node of the connection — each Kafka broker or controller, each RabbitMQ node, each Memcached node — and an incident names the node.

Three rules hold across the list. A figure BROKA could not read is unevaluatable, with the reason, never 0. A yes-or-no figure is a number, 1 or 0, so it takes the same comparisons as any other. And there are no raw counters: a threshold on a total that only ever grows would fire once and stay fired, so a counter appears as its rate or not at all.

The platforms other than Kafka are in Commercial; their metrics are offered only where the installation serves them.

Kafka

Metric	Key	Unit	Lands on
Consumer group lag	kafka.group.lag	messages	consumer group
Consumer group time lag	kafka.group.time_lag	milliseconds	consumer group
Consume rate	kafka.group.consume_rate	per second	consumer group
Consumer group members	kafka.group.members	count	consumer group
Topic produce rate	kafka.topic.produce_rate	per second	topic
Topic message count	kafka.topic.messages	messages	topic
Topic size	kafka.topic.size_bytes	bytes	topic
Time since last write	kafka.topic.last_write_age	milliseconds	topic
Partition count	kafka.topic.partitions	count	topic
Under-replicated partitions	kafka.cluster.urp	partitions	connection
Cluster messages in	kafka.cluster.messages_in_rate	per second	connection
Broker count	kafka.cluster.brokers	count	connection
Cluster size	kafka.cluster.size_bytes	bytes	connection

Topic and group figures come from the snapshot BROKA refreshes about every thirty seconds; a snapshot older than three refreshes makes them unevaluatable, with its age.

Kafka Connect

Metric	Key	Unit	Lands on
Connect cluster reachable	kafka.connect.cluster.reachable	1 / 0	Connect cluster
Connectors failed	kafka.connect.cluster.connectors_failed	count	Connect cluster
Connector tasks failed	kafka.connect.cluster.tasks_failed	count	Connect cluster
Connector failed	kafka.connect.connector.failed	1 / 0	connector
Connector paused (or stopped)	kafka.connect.connector.paused	1 / 0	connector
Connector unassigned	kafka.connect.connector.unassigned	1 / 0	connector
Connector failed tasks	kafka.connect.connector.failed_tasks	count	connector
Connector tasks running	kafka.connect.connector.running_ratio	ratio	connector

A worker that does not answer reads reachable = 0, which a rule can fire on; the other figures are then unevaluatable. A connector is shown by its Connect cluster's name and its own.

Schema Registry

Metric	Key	Unit	Lands on
Schema Registry reachable	kafka.schema_registry.reachable	1 / 0	connection
Schema Registry response time	kafka.schema_registry.response_ms	milliseconds	connection
Schema Registry read-only	kafka.schema_registry.read_only	1 / 0	connection
Schema Registry subjects	kafka.schema_registry.subjects	count	connection

On a connection with no registry these are unevaluatable and say so. Read-only is the registry's mode READONLY or IMPORT, in which a producer registering a new schema version fails.

Kafka over JMX

Read only where the connection's JMX is set up; without it, every one of them is unevaluatable with that reason. A node whose agent did not answer is unevaluatable on its own while the others are measured.

Metric	Key	Unit	Lands on
Active controllers	kafka.cluster.active_controllers	count	connection
Offline partitions	kafka.cluster.offline_partitions	partitions	connection
Unclean leader elections	kafka.cluster.unclean_elections_rate	per second	connection
Leader elections	kafka.cluster.leader_elections_rate	per second	connection
Under-min-ISR partitions	kafka.broker.under_min_isr	partitions	per node
At-min-ISR partitions	kafka.broker.at_min_isr	partitions	per node
ISR shrinks	kafka.broker.isr_shrinks_rate	per second	per node
Offline log directories	kafka.broker.offline_log_dirs	count	per node
Broker not running	kafka.broker.not_running	1 / 0	per node
Metadata lag	kafka.node.metadata_lag_ms	milliseconds	per node
Produce p99	kafka.broker.produce_p99_ms	milliseconds	per node
Consumer fetch p99	kafka.broker.fetch_consumer_p99_ms	milliseconds	per node
Follower fetch p99	kafka.broker.fetch_follower_p99_ms	milliseconds	per node
Produce mean time	kafka.broker.produce_mean_ms	milliseconds	per node
Consumer fetch mean time	kafka.broker.fetch_consumer_mean_ms	milliseconds	per node
Failed produce requests	kafka.broker.failed_produce_rate	per second	per node
Failed fetch requests	kafka.broker.failed_fetch_rate	per second	per node
Bytes rejected	kafka.broker.bytes_rejected_rate	bytes per second	per node
Request handlers idle	kafka.broker.request_handler_idle_percent	percent	per node
Network processors idle	kafka.broker.network_idle_percent	percent	per node
Request queue	kafka.broker.request_queue	count	per node
Response queue	kafka.broker.response_queue	count	per node
Produce purgatory	kafka.broker.purgatory_produce	count	per node
Fetch purgatory	kafka.broker.purgatory_fetch	count	per node
Log flush p99	kafka.broker.log_flush_p99_ms	milliseconds	per node
Heap used	kafka.node.heap_used_percent	percent	per node
GC time	kafka.node.gc_time_rate	milliseconds per second	per node
Process CPU	kafka.node.cpu_percent	percent	per node
File descriptors used	kafka.node.open_fds_percent	percent	per node
Threads	kafka.node.threads	count	per node
Broker bytes in	kafka.broker.bytes_in_rate	bytes per second	per node
Broker bytes out	kafka.broker.bytes_out_rate	bytes per second	per node
Broker messages in	kafka.broker.messages_in_rate	per second	per node

The four cluster figures are published only from a whole answer: active controllers needs every controller, the other three the one node that says it is the active controller. Leader and unclean election rates are not published by every Kafka — see JMX metrics. A consumer fetch waits up to fetch.max.wait.ms by design, so set a fetch-time threshold above it. Broker messages in is per broker only; the cluster's figure is Cluster messages in, from offsets. A topic's own JMX rates are shown on its page and are not alertable.

Redis

Metric	Key	Unit	Lands on
Memory used	redis.memory.used_bytes	bytes	per node
Memory used (of maxmemory)	redis.memory.used_percent	percent	connection
Memory fragmentation ratio	redis.memory.fragmentation_ratio	ratio	connection
Keyspace hit rate	redis.hit_rate	ratio	connection
Operations per second	redis.ops_per_sec	per second	per node
Connected clients	redis.clients.connected	count	per node
Blocked clients	redis.clients.blocked	count	connection
Key count	redis.keys	count	connection
Replication lag	redis.replication.lag_bytes	bytes	connection
Time since last replication I/O	redis.replication.last_io_age	seconds	connection
Stream length	redis.stream.length	entries	stream
Stream group lag	redis.stream.group_lag	entries	stream
Pending entries	redis.stream.pending	entries	stream

On a Redis Cluster the figures a single node answers for are measured per node and name the shard; the ones that describe one node rather than the deployment are unevaluatable there. Memory used (of maxmemory) is unevaluatable on a Redis with no maxmemory.

RabbitMQ

Metric	Key	Unit	Lands on
Queue depth	rabbitmq.queue.messages	messages	queue
Ready messages	rabbitmq.queue.ready	messages	queue
Unacknowledged messages	rabbitmq.queue.unacknowledged	messages	queue
Queue consumers	rabbitmq.queue.consumers	count	queue
Consumer utilisation	rabbitmq.queue.consumer_utilisation	ratio	queue
Redelivery rate	rabbitmq.queue.redeliver_rate	per second	queue
Queue memory	rabbitmq.queue.memory_bytes	bytes	queue
Total messages	rabbitmq.cluster.messages	messages	connection
Total unacknowledged	rabbitmq.cluster.unacknowledged	messages	connection
Publish rate	rabbitmq.cluster.publish_rate	per second	connection
Deliver rate	rabbitmq.cluster.deliver_rate	per second	connection
Unroutable rate	rabbitmq.cluster.unroutable_rate	per second	connection
Connections	rabbitmq.cluster.connections	count	connection
Consumers	rabbitmq.cluster.consumers	count	connection
Node memory used	rabbitmq.node.memory_used_percent	percent	per node
Node free disk	rabbitmq.node.disk_free_bytes	bytes	per node
Node file descriptors used	rabbitmq.node.file_descriptors_used_percent	percent	per node
Node alarm	rabbitmq.node.alarm	1 / 0	per node
Node running	rabbitmq.node.running	1 / 0	per node
Shovel running	rabbitmq.shovel.running	1 / 0	shovel
Shovel blocked	rabbitmq.shovel.blocked	1 / 0	shovel
Shovel pending	rabbitmq.shovel.pending	messages	shovel
Shovels not running	rabbitmq.cluster.shovels_not_running	count	connection
Federation link running	rabbitmq.federation_link.running	1 / 0	federation link
Federation links not running	rabbitmq.cluster.federation_links_not_running	count	connection

The rates are unevaluatable where the management plugin's statistics are off. A shovel that is starting, terminated or reported with no state reads running = 0; a federation link that is starting, in error or shut down likewise.

Apache Artemis

Metric	Key	Unit	Lands on
Queue message count	artemis.queue.message_count	messages	queue
Queue consumers	artemis.queue.consumer_count	count	queue
Delivering	artemis.queue.delivering_count	messages	queue
Scheduled	artemis.queue.scheduled_count	messages	queue
Queue paused	artemis.queue.paused	1 / 0	queue
Oldest message age	artemis.queue.first_message_age	milliseconds	queue
Unrouted messages	artemis.address.unrouted_count	messages	address
Address size	artemis.address.size_bytes	bytes	address
Address limit used	artemis.address.limit_percent	percent	address
Address paging	artemis.address.paging	1 / 0	address
Broker live	artemis.broker.live	1 / 0	connection
Bridge connected	artemis.bridge.connected	1 / 0	bridge
Bridge pending acknowledgements	artemis.bridge.pending_ack	messages	bridge
Broker connection connected	artemis.broker_connection.connected	1 / 0	broker connection
Cluster members	artemis.cluster_connection.nodes	count	connection

Cluster members counts the nodes this broker's cluster connections see, and is unevaluatable on a broker with none. Connected means the link is up, not that the two brokers are in step.

Memcached

Metric	Key	Unit	Lands on
Node memory used	memcached.node.memory_used_percent	percent	per node
Node items	memcached.node.items	count	per node
Node connections	memcached.node.connections	count	per node
Node connections used	memcached.node.connections_percent	percent	per node
Node hit rate	memcached.node.hit_rate	ratio	per node
Node unreachable	memcached.node.unreachable	1 / 0	per node
Node eviction rate	memcached.node.eviction_rate	per second	per node

A node evicts before memory used reaches 100%, because items are stored in fixed-size chunks; the eviction rate is the direct signal. It needs two readings, so it is unevaluatable on a node's first reading and after a restart.

What a rule cannot express

No trend — "lag is rising" — because BROKA keeps no time series; a rule's hold for window is the substitute, a breach seen to continue rather than a slope. No learned thresholds, for the same reason. And one metric per rule: "depth above 1,000 and no consumers" is two rules on the same channel.