StatsD metric instrumentation: push metrics over UDP, choose the right metric type, name consistently. Invoke whenever task involves any interaction with StatsD metrics — emitting counters, gauges, timers, histograms, sets, or distributions; configuring DogStatsD tags; tuning sampling rates; or integrating with Graphite, Prometheus, or Telegraf backends.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
StatsD metric instrumentation: push metrics over UDP, choose the right metric type, name consistently. Invoke whenever task involves any interaction with StatsD metrics — emitting counters, gauges, timers, histograms, sets, or distributions; configuring DogStatsD tags; tuning sampling rates; or integrating with Graphite, Prometheus, or Telegraf backends.
StatsD
Choose the right metric type, name it with dot-delimited hierarchy, tag dimensions instead of encoding them in names.
StatsD is fire-and-forget: UDP means zero latency impact, but wrong metric types or bad naming corrupt your data
silently.
References
Metric types — [${CLAUDE_SKILL_DIR}/references/metric-types.md]: Wire format details, type comparison, sampling
correction
Naming — [${CLAUDE_SKILL_DIR}/references/naming.md]: Graphite namespace mapping, character rules, naming
examples
DogStatsD — [${CLAUDE_SKILL_DIR}/references/dogstatsd.md]: Events format, service checks, protocol versions,
distributions vs histograms
Wrong metric type = wrong math at the server. A gauge used as a counter loses data between flushes; a counter used as a
gauge produces meaningless rates.
Counter (|c)
Measures rate of events over time. Server sums all values during flush interval, resets to 0 after flush, reports both
raw count and per-second rate.
Sample rate correction: value multiplied by 1/rate
Supports sampling (|@<rate>)
Gauge (|g)
Instantaneous value at a point in time. Server stores last value received, retains between flushes (sticky).
Use for: queue depth, active connections, memory/CPU usage, thread pool size
Signed values (+N, -N) modify current value incrementally
Cannot set to a negative number directly — set to 0 first, then decrement
Do not sample gauges — server cannot correct for sampling on point-in-time values
Timer (|ms)
Duration of an operation in milliseconds. Server computes per flush interval: count, mean, upper (max), lower (min),
sum, stddev, median, configurable percentiles (p90, p95, p99).
Use for: HTTP request latency, DB query duration, function execution time
Supports sampling (|@<rate>)
Histogram (|h)
Distribution of values over time. Identical to timer in most implementations. DogStatsD treats histograms as the native
distribution type.
Use for: request payload sizes, response body sizes, batch sizes
Always namespace by service name — myapp.api.requests not just requests
Use dot-delimited hierarchy
Include the unit: .ms, .bytes, .total, .items
Dimensions go in tags, not metric names (when tags are available)
Use lowercase everywhere — some backends are case-sensitive
Use underscores within path segments: http_request not httpRequest
No dashes — they break Graphite navigation
See ${CLAUDE_SKILL_DIR}/references/naming.md for Graphite namespace mapping, character rules table, and naming
anti-patterns.
Tags (DogStatsD)
Format: metric.name:1|c|#key1:value1,key2:value2 — comma-separated, no spaces.
Tag Rules
Use tags for dimensions you will filter or group by — not metric names
Keep cardinality bounded — each unique tag combination creates a separate time series
No spaces in tag values — use underscores: region:us_east
Unified Service Tagging
Set these as global/constant tags on the client — attach to every metric automatically:
env — Deployment environment (env:production)
service — Service name (service:payment-api)
version — Deployed version (version:2.1.0)
Tag Cardinality
Rule of thumb: if a tag can have >1000 distinct values, do not use it. Use logs or traces for high-cardinality data.
env:production — ~3-5 values: Yes
method:GET — ~7 values: Yes
status_code:200 — ~20-50 values: Yes
endpoint:/api/users — ~50-200 values: Caution
user_id:12345 — Unbounded: No
Aggregation and Flush
The flush cycle determines metric resolution. Default: 10 seconds.
Counters reset to 0 after flush; gauges are sticky (retain last value)
If no counter values received during flush: behavior depends on deleteCounters config (default: send 0)
Enable client-side aggregation for high-throughput applications (Go v5.0+, Java v3.0+, .NET v7.0+) — pre-aggregates
before sending to Agent
See ${CLAUDE_SKILL_DIR}/references/aggregation.md for flush mechanics, Graphite downsampling rules, DogStatsD
aggregation details, and pre-aggregated timestamps.
Client Patterns
Initialization
One client instance per application — do not create per-request
Set namespace prefix — auto-prepends to all metric names
Set global/constant tags — env, service, version set once
Close/flush on shutdown — buffered metrics lost otherwise
Buffering
Enable client-side buffering — packs multiple metrics into single UDP packets. Reduces syscall overhead in hot paths.
Most modern DogStatsD clients buffer by default. Call flush() before shutdown.
Sampling
Client randomly decides whether to send each metric based on sample rate. Datagram includes |@<rate> so server
corrects the count.
< 1000 metrics/sec — rate=1.0 (no sampling)
1000-10000/sec — rate=0.5 to 0.1 for counters/timers
> 10000/sec — rate=0.1 or lower; enable client-side aggregation
Never sample gauges or sets — server cannot correct for these types.
See ${CLAUDE_SKILL_DIR}/references/client-patterns.md for high-throughput tuning steps, error handling, and Kubernetes
deployment patterns.
