Observability

Each shard owner keeps its own counters. Updates are ordinary local integer operations; the hot path does not use atomic counters or a global metrics lock.

Query a shard with:

redis-cli -p 6379 INFO

INFO section returns one block; the name compares without case or spaces (INFO commandcounters), all, everything and default return every block, and an unknown name returns an empty reply. INFO prometheus returns the Prometheus text below. Every field INFO writes is in this table, and a test fails when one is added without a row:

Section	Field	Meaning
Server	ruvio_version	Server version
Server	ruvio_backend	Runtime backend (evented)
Server	ruvio_health	ok while the shard answers
Server	ruvio_readiness	ready once recovery finished
Server	ruvio_durability	Durability profile
Server	ruvio_persistence_state	ok, or error after a WAL or sync failure
Server	recovered_wal_records	WAL records replayed at startup
Server	wal_last_sequence	Last WAL sequence written
Server	wal_unsynced_bytes	WAL bytes written but not yet synced
Server	ruvio_shard_id	This shard
Server	ruvio_shard_count	Shards in the process
Stats	connected_clients	Open client connections on this shard
Stats	total_connections_received	Accepted connections, process-wide
Stats	rejected_connections	Connections refused by the connection limits, process-wide
Stats	total_net_input_bytes	Bytes read from this shard’s client connections
Stats	total_net_output_bytes	Bytes written to this shard’s client connections
Buffers	input_buffer_peak_bytes	Largest connection input buffer seen
Buffers	output_buffer_peak_bytes	Largest connection output buffer seen
Buffers	pubsub_queue_peak_frames	Longest subscriber queue seen, in frames
Buffers	pubsub_queue_peak_bytes	Longest subscriber queue seen, in bytes
Memory	charged_memory_bytes	Memory charged against maxmemory
Keyspace	keys	Keys in every database
Keyspace	expired_keys	Keys removed at their deadline
Keyspace	expired_keys_passive	Of those, found on lookup or write
Keyspace	expired_keys_active	Of those, found by the bounded expire cycle
Keyspace	evicted_keys	Keys removed by the eviction policy
Keyspace	expires	Keys that carry a deadline, overdue ones included
Keyspace	db<N>	keys=…,expires=…,avg_ttl=… per non-empty database, as in Redis
Command counters	total_commands_processed	Commands run
Command counters	get_commands	GET calls
Command counters	set_commands	SET calls
Command counters	increment_commands	INCR, DECR, INCRBY, DECRBY calls
Command counters	other_commands	Every other command
Memory breakdown	key_memory_bytes	Charged key bytes
Memory breakdown	value_memory_bytes	Charged value bytes
Memory breakdown	collection_memory_bytes	Charged collection bytes
Memory breakdown	index_memory_bytes	Primary index bytes
Memory breakdown	pooled_value_bytes	Bytes kept in the value pool for reuse
Memory breakdown	persistence_memory_bytes	WAL and snapshot buffers
Memory breakdown	lazyfree_pending_bytes	Bytes waiting for the lazy-free thread
Snapshot	snapshot_in_progress	1 while a snapshot walk is open
Snapshot	snapshot_journal_bytes	Bytes the open snapshot keeps for keys written during it
Snapshot	snapshot_aborts_total	Snapshots given up
Snapshot	snapshot_overload_skips_total	Checkpoints skipped because the shard was busy
Snapshot	rdb_bgsave_in_progress	1 while a BGSAVE runs
Latencystats	latency_percentiles_usec_<command>	p50, p99, p99.9 and samples; only with latency_histograms on
Logs	log_pages	Value-log pages in use
Logs	log_extents	Value-log extents held
Logs	log_free_bytes	Free bytes inside held pages
Logs	log_pages_released_total	Pages given back
Logs	log_relocated_bytes_total	Bytes moved to compact pages
Slabs	slab_pages	Slab pages in use
Slabs	slab_records	Records stored in slabs
Slabs	slab_class_<N>b	Pages and records of one slot size, for classes in use
Replication	replication_role	off, primary or replica
Replication	wal_durable_sequence	Last WAL sequence synced to disk
Replication	replication_replicas	Connected replica sockets on a primary, otherwise 0
Replication	replication_replica_sequence	Oldest sequence written to a connected replica socket, otherwise 0
Replication	replication_send_lag	wal_durable_sequence minus that oldest send; 0 means the bytes left the primary
Replication	replica_retain_dropped	WAL holds for slow replicas dropped
Replication	master_link_status	Replica only: up or down
Replication	master_last_io_seconds_ago	Replica only: seconds since the last frame, -1 before the first
Replication	slave_repl_offset	Replica only: last applied WAL sequence
Replication	master_sync_in_progress	Replica only: 1 while a full resync is requested
Cluster	cluster_enabled	1 in cluster mode

The replica’s own wal_durable_sequence is the copy you can promote.

In multi-shard mode, query every consecutive shard port. Counters are intentionally not aggregated in the server because aggregation would introduce cross-core synchronization. Monitoring systems should scrape and sum them externally.

Prometheus text is available without adding hot-path atomics or a global lock:

redis-cli INFO prometheus

Optional HTTP scrape binds a separate listener. Set metrics_address / RUVIO_METRICS_ADDR / --metrics-address. Multi-shard processes increment that port by shard id (9121, 9122, …). A scrape is a control-plane read on the owning shard; it does not update command counters.

curl http://127.0.0.1:9121/metrics

Expiration and eviction counters live on the owning shard. Linux updates them as ordinary local integers; portable scrapes read the same shard fields. A scrape does not run the active expire cycle or evict keys.

Prometheus text includes:

• ruvio_net_input_bytes_total and ruvio_net_output_bytes_total, per shard
• ruvio_connections_total, ruvio_rejected_connections_total (refused by the connection limits), ruvio_rate_limited_commands_total (refused by the per-connection token bucket) and ruvio_protocol_errors_total (frames the parser could not read, which close the connection). These four are process-wide and reported on shard 0 only.
• ruvio_expired_keys_total{reason="passive|active"}
• ruvio_keys_with_ttl
• ruvio_evicted_keys_total
• ruvio_eviction_scans_total / ruvio_eviction_slots_inspected_total (victim-search work, including empty hash buckets and searches with no victim)
• ruvio_allocation_events_total / ruvio_allocation_bytes_total
• pool retention remains ruvio_pooled_value_bytes

Charged memory is a deterministic storage-budget model, not process RSS. See Memory Accounting for category definitions and limitations.