Documentation
¶
Overview ¶
Package obs holds the observability primitives: structured audit log middleware and Prometheus metrics.
Index ¶
- func NewAuditMiddleware(w io.Writer) func(http.Handler) http.Handler
- type InventoryCollector
- type InventorySource
- type JobMetrics
- func (m *JobMetrics) ChildFailure(kind string)
- func (m *JobMetrics) JobEnqueued(kind string)
- func (m *JobMetrics) JobFinished(kind, result string)
- func (m *JobMetrics) JobStarted(kind string)
- func (m *JobMetrics) ObserveDuration(kind string, d time.Duration)
- func (m *JobMetrics) Rollback(kind string)
- type Metrics
- type PruneMetrics
- type StatsCollector
- type StatsSource
- type VolumeUsageCollector
- type VolumeUsageSource
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
Types ¶
type InventoryCollector ¶ added in v1.0.23
type InventoryCollector struct {
// contains filtered or unexported fields
}
InventoryCollector renders per-container and per-host state from the warm inventory cache at scrape time.
It is a collector rather than a set of pushed gauges for two reasons: it does no podman I/O, so a scrape can never block behind an unreachable host; and it renders present state, so a deleted or renamed instance's series disappear on the next poll instead of freezing at their last value forever.
func NewInventoryCollector ¶ added in v1.0.23
func NewInventoryCollector(reg prometheus.Registerer, src InventorySource, hosts func() []string) *InventoryCollector
NewInventoryCollector builds the collector and registers it on reg.
hosts must be the same host-list function the inventory poller uses. A host that has never been polled then still reports podman_api_host_reachable 0 rather than vanishing from the metric set, which would read as "fine".
func (*InventoryCollector) Collect ¶ added in v1.0.23
func (c *InventoryCollector) Collect(ch chan<- prometheus.Metric)
func (*InventoryCollector) Describe ¶ added in v1.0.23
func (c *InventoryCollector) Describe(ch chan<- *prometheus.Desc)
type InventorySource ¶ added in v1.0.23
type InventorySource interface {
InventorySnapshot() map[string]instance.HostInventory
}
InventorySource supplies the current warm-cache inventory snapshot. Implemented by *instance.Service.
type JobMetrics ¶
type JobMetrics struct {
// contains filtered or unexported fields
}
JobMetrics implements jobs.Metrics with Prometheus collectors. Created with an explicit Registerer so production registers on the default registry (NewJobMetrics(prometheus.DefaultRegisterer)) and tests use a private one.
func NewJobMetrics ¶
func NewJobMetrics(reg prometheus.Registerer) *JobMetrics
NewJobMetrics builds and registers the collectors on reg.
func (*JobMetrics) ChildFailure ¶
func (m *JobMetrics) ChildFailure(kind string)
ChildFailure records an evacuate child migration failure.
func (*JobMetrics) JobEnqueued ¶
func (m *JobMetrics) JobEnqueued(kind string)
JobEnqueued records a new job being enqueued.
func (*JobMetrics) JobFinished ¶
func (m *JobMetrics) JobFinished(kind, result string)
JobFinished records a terminal job outcome (succeeded/failed/canceled).
func (*JobMetrics) JobStarted ¶
func (m *JobMetrics) JobStarted(kind string)
JobStarted records a job being claimed and started.
When JobEnqueued is wired at dispatch sites, restore m.enqueued.Dec() here so the queue-depth gauge increments on enqueue and decrements on start.
func (*JobMetrics) ObserveDuration ¶
func (m *JobMetrics) ObserveDuration(kind string, d time.Duration)
ObserveDuration records a job handler's wall-clock duration.
func (*JobMetrics) Rollback ¶
func (m *JobMetrics) Rollback(kind string)
Rollback records a migrate rollback event.
type Metrics ¶
type Metrics struct {
// contains filtered or unexported fields
}
Metrics holds the Prometheus collectors. New() registers them with the default registry; call Handler() to expose /metrics.
func (*Metrics) Middleware ¶
Middleware returns an HTTP middleware that records request count and duration. The route label uses r.Pattern (the registered mux pattern, available in Go 1.22+) to avoid label cardinality explosion from path parameters. Falls back to "_other" when the pattern is not populated.
type PruneMetrics ¶
type PruneMetrics struct {
// contains filtered or unexported fields
}
PruneMetrics implements the prune.Metrics interface (structurally — obs must not import prune) with Prometheus counters. Created with an explicit Registerer so production registers on the default registry (NewPruneMetrics(prometheus.DefaultRegisterer)) and tests use a private one.
func NewPruneMetrics ¶
func NewPruneMetrics(reg prometheus.Registerer) *PruneMetrics
NewPruneMetrics builds and registers the prune collectors on reg.
func (*PruneMetrics) Reclaimed ¶
func (m *PruneMetrics) Reclaimed(host, scope string, bytes int64)
Reclaimed adds reclaimed bytes for a scope.
func (*PruneMetrics) RunDone ¶
func (m *PruneMetrics) RunDone(host, result string)
RunDone records one finished prune run.
type StatsCollector ¶ added in v1.0.28
type StatsCollector struct {
// contains filtered or unexported fields
}
StatsCollector renders per-container resource samples at scrape time.
Like InventoryCollector it does no podman I/O — it reads two cached snapshots. The inventory snapshot is what supplies template/slug: podman reports stats by container name only, so a sample whose container is not in the inventory is dropped rather than emitted under a guessed identity.
func NewStatsCollector ¶ added in v1.0.28
func NewStatsCollector(reg prometheus.Registerer, stats StatsSource, inv InventorySource, hosts func() []string) *StatsCollector
NewStatsCollector builds the collector and registers it on reg.
hosts must be the same host-list function the inventory poller and InventoryCollector use, and it — not the cache maps — is what Collect enumerates. Neither cache is pruned when a host leaves hosts/*.yaml on SIGHUP, so iterating a snapshot would re-emit that host's last sample on every scrape forever: cumulative counters frozen mid-flight read as idle containers, and the documented mitigation (gate on podman_api_host_reachable == 1) is unavailable because InventoryCollector has already stopped emitting that series for the departed host.
func (*StatsCollector) Collect ¶ added in v1.0.28
func (c *StatsCollector) Collect(ch chan<- prometheus.Metric)
func (*StatsCollector) Describe ¶ added in v1.0.28
func (c *StatsCollector) Describe(ch chan<- *prometheus.Desc)
type StatsSource ¶ added in v1.0.28
StatsSource supplies the latest per-host container resource samples. Implemented by *instance.Service.
type VolumeUsageCollector ¶ added in v1.0.28
type VolumeUsageCollector struct {
// contains filtered or unexported fields
}
VolumeUsageCollector renders volume sizes at scrape time, attributing each volume to its instance via the inventory snapshot. Does no podman I/O.
func NewVolumeUsageCollector ¶ added in v1.0.28
func NewVolumeUsageCollector(reg prometheus.Registerer, src VolumeUsageSource, inv InventorySource, hosts func() []string) *VolumeUsageCollector
NewVolumeUsageCollector builds the collector and registers it on reg.
hosts must be the same host-list function the inventory poller and InventoryCollector use, and it — not the cache map — is what Collect enumerates. volumeUsageCache deliberately keeps a host's last sizing through a failed walk and has no drop path at all, so iterating a snapshot would leave a host removed on SIGHUP reporting frozen sizes and an ever-climbing podman_api_volume_usage_age_seconds forever.
func (*VolumeUsageCollector) Collect ¶ added in v1.0.28
func (c *VolumeUsageCollector) Collect(ch chan<- prometheus.Metric)
func (*VolumeUsageCollector) Describe ¶ added in v1.0.28
func (c *VolumeUsageCollector) Describe(ch chan<- *prometheus.Desc)
type VolumeUsageSource ¶ added in v1.0.28
type VolumeUsageSource interface {
VolumeUsageSnapshot() map[string]instance.HostVolumeUsage
}
VolumeUsageSource supplies the latest per-host volume sizing. Implemented by *instance.Service.
Source Files