Metrics Reference

All metrics are gauges named pflex_<object>_<metric> and carry a cluster label (the configured cluster name) plus cluster_id (the PowerFlex System id).

This page documents metric families and conventions. The authoritative, concrete emitted metric-name set is generated from the collector via the dump test (PFLEX_DUMP_METRICS=1 go test ./internal/powerflex/ -run TestDumpEmittedMetricNames), not hand-enumerated here.

Object prefixes

PowerFlex type	Metric prefix
System (cluster)	`pflex_cluster_`
SDS	`pflex_sds_`
SDC	`pflex_sdc_`
Volume	`pflex_volume_`
StoragePool	`pflex_storagepool_`
Device	`pflex_device_`
ProtectionDomain	`pflex_protectiondomain_`

Identity & parent labels

Beyond cluster / cluster_id, each object type adds identity and parent labels, resolved from the PowerFlex relationship graph:

Type	Additional labels
SDS	`sds`, `sds_id`, `protection_domain_name`, `protection_domain_id`
SDC	`sdc_name`, `sdc_id` (name falls back to the SDC IP, then id)
Volume	`volume_name`, `volume_id`, `storage_pool_name`, `storage_pool_id`, `protection_domain_name`, `protection_domain_id`
StoragePool	`storage_pool_name`, `storage_pool_id`, `protection_domain_name`, `protection_domain_id`
Device	`device_name`, `device_id`, `device_path`, `sds`, `sds_id`, `storage_pool_name`, `storage_pool_id`, `protection_domain_name`, `protection_domain_id`
ProtectionDomain	`protection_domain_name`, `protection_domain_id`

Kubernetes enrichment labels (optional)

When Kubernetes workload enrichment is enabled, two object types gain extra labels mapped from the cluster's PVs and Nodes:

Type	Added labels
Volume	`namespace`, `persistent_volume_claim`, `persistent_volume`, `storage_class`
SDC	`k8s_node`

These keys are present on every pflex_volume_* / pflex_sdc_* series while enrichment is on (empty when a workload can't be resolved), so the label-key set stays consistent. They let you aggregate storage performance by namespace, claim, or node — e.g. sum by (namespace) (pflex_volume_iops).

Derived performance metrics

PowerFlex reports *Bwc accumulators (numOccured, numSeconds, totalWeightInKb). Each becomes three gauges, plus *Latency accumulators become one:

Metric	Derivation	Extra labels
`pflex_<obj>_iops`	`numOccured / numSeconds` (per second)	`op`, `direction`
`pflex_<obj>_bandwidth_kb_per_second`	`totalWeightInKb / numSeconds` (KB/s)	`op`, `direction`
`pflex_<obj>_io_size_kb`	`totalWeightInKb / numOccured` (avg KB)	`op`, `direction`
`pflex_<obj>_latency`	`totalWeightInKb / numOccured`	`op`, `direction`

direction is read or write (or empty for accumulators with neither).
op is the accumulator prefix: primary, total, fwdRebuild, bckRebuild, normRebuild, rebalance, volMigration, userData, userDataSdc, …

Do not use rate()

iops and bandwidth_kb_per_second are already per-second values computed by PowerFlex. In PromQL aggregate them with sum / avg by labels — never wrap them in rate() or you will double-rate.

Scalar metrics

Every non-accumulator statistic becomes a scalar gauge named from the API field in snake_case, e.g.:

pflex_cluster_max_capacity_in_kb, pflex_cluster_capacity_in_use_in_kb, pflex_cluster_unused_capacity_in_kb, pflex_cluster_spare_capacity_in_kb, pflex_cluster_compression_ratio, …
pflex_storagepool_max_capacity_in_kb, pflex_storagepool_spare_capacity_in_kb, …
pflex_sdc_num_of_mapped_volumes
pflex_device_avg_read_size_in_bytes, pflex_device_avg_read_latency_in_microsec, …

Capacity values are reported in KiB (*_in_kb), matching the PowerFlex API. Multiply by 1024 in PromQL if you need bytes.

Redundancy & rebuild/rebalance (Gen1; System / StoragePool / ProtectionDomain):

Capacity-at-risk: pflex_<obj>_degraded_failed_capacity_in_kb, pflex_<obj>_failed_capacity_in_kb, pflex_<obj>_degraded_healthy_capacity_in_kb, pflex_<obj>_spare_capacity_in_kb.
Rebuild/rebalance remaining capacity: pflex_<obj>_fwd_rebuild_capacity_in_kb, pflex_<obj>_bck_rebuild_capacity_in_kb, pflex_<obj>_norm_rebuild_capacity_in_kb, pflex_<obj>_rebalance_capacity_in_kb (the rate of rebuild/rebalance remains the *RebuildBwc-derived iops/bandwidth).
Data-movement job counts: pflex_<obj>_pending_moving_in_bck_rebuild_jobs, pflex_<obj>_active_moving_in_bck_rebuild_jobs, and the out/fwd_rebuild/rebalance variants.
Snapshot capacity: pflex_<obj>_snapshot_capacity_in_kb, pflex_<obj>_net_snapshot_capacity_in_kb, pflex_<obj>_snap_capacity_in_use_in_kb.

Back-end latency (Gen1): pflex_protectiondomain_latency{op="target"|"journaler"} (and op="target" on StoragePool) split read/write — complements the host-side op="userData" latency for localizing a latency problem to the device/journal tier.

Inventory counts (Gen1; from System instance properties): pflex_cluster_num_of_volumes, pflex_cluster_num_of_sds, pflex_cluster_num_of_devices.

Gen2 differences

The generation is detected per cluster (storage-pool dataLayout: ErasureCoding = Gen2) and exposed as pflex_cluster_generation{generation="gen2"}. Gen2 collects from the v5 metrics API, where values are pre-computed, so the shapes are the same but with explicit units and Gen2-specific labels:

Object prefixes: Gen1 pflex_sds_* becomes pflex_storagenode_* (SDS was renamed StorageNode). Gen2 adds pflex_devicegroup_* (with a media_type label, incl. PMEM/WRC op values) and pflex_sdt_* (NVMe/TCP path latency). System, Volume, StoragePool, Device, Sdc and ProtectionDomain keep their prefixes.
Unit-explicit derived names (Gen2 values are bytes/s, bytes, µs): pflex_<obj>_iops (shared), pflex_<obj>_bandwidth_bytes_per_second, pflex_<obj>_io_size_bytes, pflex_<obj>_latency_microseconds.
op values: host, device, device_local, device_remote, storage_fe, device_pmem, wrc, rebuild, rebalance, controller_to_host, host_to_controller. direction is read/write/trim/"".
Scalars are bytes/ratios named from the v5 field: pflex_<obj>_physical_used, _raw_total, _logical_provisioned, _compression_ratio, _data_reduction_ratio, _utilization_ratio, etc. (distinct from Gen1's *_in_kb).
Volume carries volume_type (BaseVolume/ThinClone) and Device carries storage_node_name/id. Both Gen1 and Gen2 forms of these two metrics share a union label-key set (inapplicable keys empty) so a single exporter can serve mixed-generation fleets and keep /metrics valid.

Health & meta metrics

Metric	Labels	Meaning
`pflex_exporter_build_info`	`version`, `goversion`	Always `1`; an exporter-level metric whose `version` (ldflag-set build version) and `goversion` labels identify the running build. Unlike `pflex_*` metrics it carries no `cluster`/identity label — it describes the exporter process, not a cluster.
`pflex_up`	`cluster`	`1` if the cluster was scraped successfully this cycle, else `0`.
`pflex_last_scrape_timestamp_seconds`	`cluster`	Unix time of the last successful collection.
`pflex_cluster_generation`	`cluster`, `generation`	Always `1`; the `generation` label is `gen1`, `gen2`, or `unknown`.
`pflex_<obj>_health`	object identity/parent labels	Operational severity: `0`=healthy, `1`=degraded, `2`=failed/disconnected/unknown. Emitted for SDS/StorageNode, Device, SDC, and Sdt (Gen2 only). An unreported (empty) optional state field carries no signal and does not force `2`.
`pflex_<obj>_info`	identity labels + raw state strings	Always `1`; carries raw PowerFlex state strings (`mdm_connection_state`, `membership_state`, `maintenance_state`, `device_state`; Device also `temperature_state`, `ssd_end_of_life_state`, `error_state`; Sdt `sdt_state`).
`pflex_volume_mapped_sdc`	volume identity/parent labels + `sdc_id`, `sdc_ip`	Always `1`; one series per volume→SDC mapping, correlating a volume with each host consuming it.

Operational state

State gauges are derived from object properties in GET /api/instances (not from the statistics API). The *_health value is the worst severity across the object's state fields; the matching *_info metric preserves the raw strings as labels. A missing or unrecognized state maps to 2 (unknown), so a lost signal is surfaced rather than hidden.

Expected state fields are present on PowerFlex 4.5+ (Gen1) and 5.x (Gen2). Alert on pflex_<obj>_health > 0; join to a host with pflex_volume_mapped_sdc (e.g. pflex_volume_mapped_sdc * on(volume_id) group_left() pflex_volume_iops, which yields one result per volume↔SDC pair carrying the volume's IOPS).

An unknown-generation cluster (no recognizable storage-pool layout) reports pflex_up=1 and pflex_cluster_generation{generation="unknown"}=1 and falls back to the Gen1 collection path.