metrics
Public
Pull mirror status: ok
K
Merge pull request #24 from InjectiveLabs/statsd-refresh-stale-socket
pflag
pflag helpers
.gitignore
[Feature] add raw report func
client_test.go
fix: publish a new client only once Init cannot fail, and retire loops at once
client.go
fix: publish a new client only once Init cannot fail, and retire loops at once
go.mod
Revert "Merge pull request #23 from InjectiveLabs/f/v2"
go.sum
Revert "Merge pull request #23 from InjectiveLabs/f/v2"
metrics_test.go
feat: expose Tags as maps
metrics.go
fix: bind each refresh loop to the Init that started it
otel_test.go
Revert "Merge pull request #23 from InjectiveLabs/f/v2"
otel.go
Revert "Merge pull request #23 from InjectiveLabs/f/v2"
README.md
fix: publish a new client only once Init cannot fail, and retire loops at once
reports.go
feat: add histogram helper
service_config_test.go
have a service config
service_config.go
have a service config
telegraf.go
initial commit
trace.go
feat: mark traces with tasks

Metrics

Motivation

  • We are going to have a lot of microservices requiring metrics -> Unify how + what metrics are sent

  • Maintenance in one place, only need update this lib in services implementation

Supports:

  • Telegraf

  • DataDog

  • OpenTelemetry (OTEL)

Statsd client refresh

The Telegraf and DataDog statters send over UDP. They resolve their address and dial once, at construction, and then hold that socket for the life of the process — and nothing in the write path can tell that the socket has stopped being delivered, because a connectionless send to a dead peer still succeeds locally. A process that outlives the agent it was pointed at therefore keeps "reporting" into a hole, silently, with no errors and a healthy-looking service.

To prevent that, the UDP client is rebuilt every DefaultRefreshInterval (5 minutes). A rebuild is a UDP dial with no handshake, and the client it replaces is closed, which flushes whatever it had buffered.

  • InitService starts the refresh automatically. Tune it with ServiceConfig.RefreshInterval, or set that to a negative value to switch it off.
  • Callers that use Init directly should call StartRefresh(ctx, addr, prefix, 0) once afterwards.
  • A later Init retires the loop the previous one started, immediately rather than at that loop's next tick, so a re-init cannot leave a refresh rebuilding for the address and tag format of the config it replaced.
  • An Init that fails leaves the previous client and config in place; nothing is published until every step that can fail has succeeded.

OTEL is deliberately left alone: it exports over gRPC/HTTP, which redials on its own.

Acknowledgement