Go statsd client library with zero allocation overhead, great performance and automatic reconnects.
Client has zero memory allocation per metric sent:
- ring of buffers, each buffer is UDP packet
- buffer is taken from the pool, filled with metrics, passed on to the network delivery and returned to the pool
- buffer is flushed either when it is full or when flush period comes (e.g. every 100ms)
- separate goroutines handle network operations: sending UDP packets and reconnecting UDP socket
- when metric is serialized, zero allocation operations are used to avoid
reflect
and temporary buffers
As metrics could be sent by the application at very high rate (e.g. hundreds of metrics per one request),
it is important that sending metrics doesn't cause any additional GC or CPU pressure. go-statsd
is using
buffer pools and it tries to avoid allocations while building statsd packets.
With modern container-based platforms with dynamic DNS statsd server might change its address when container
gets rescheduled. As statsd packets are delivered over UDP, there's no easy way for the client to figure out
that packets are going nowhere. go-statsd
supports configurable reconnect interval which forces DNS resolution.
While client is reconnecting, metrics are still processed and buffered.
When buffer pool is exhausted, go-statsd
starts dropping packets. Number of dropped packets is reported via
Client.GetLostPackets()
and every minute logged using log.Printf()
. Usually packets should never be dropped,
if that happens it's usually signal of enormous metric volume.
Any statsd-compatible server should work well with go-statsd
, statsite works
exceptionally well as it has great performance and low memory footprint even with huge number of metrics.
Initialize client instance with options, one client per application is usually enough:
client := statsd.NewClient("localhost:8125",
statsd.MaxPacketSize(1400),
statsd.MetricPrefix("web."))
Send metrics as events happen in the application, metrics will be packed together and delivered to statsd server:
start := time.Now()
client.Incr("requests.http", 1)
// ...
client.PrecisionTiming("requests.route.api.latency", time.Since(start))
Shutdown client during application shutdown to flush all the pending metrics:
client.Close()
Metrics could be tagged to support aggregation on TSDB side. go-statsd supports tags in InfluxDB , Datadog and Graphite formats. Format and default tags (applied to every metric) are passed as options to the client initialization:
client := statsd.NewClient("localhost:8125",
statsd.TagStyle(TagFormatDatadog),
statsd.DefaultTags(statsd.StringTag("app", "billing")))
For every metric sent, tags could be added as the last argument(s) to the function call:
client.Incr("request", 1,
statsd.StringTag("procotol", "http"), statsd.IntTag("port", 80))
Benchmark comparing several clients:
- https://github.com/alexcesaro/statsd/ (
Alexcesaro
) - this client (
GoStatsd
) - https://github.com/cactus/go-statsd-client (
Cactus
) - https://github.com/peterbourgon/g2s (
G2s
) - https://github.com/quipo/statsd (
Quipo
) - https://github.com/Unix4ever/statsd (
Unix4ever
)
Benchmark results:
BenchmarkAlexcesaro-12 5000000 333 ns/op 0 B/op 0 allocs/op
BenchmarkGoStatsd-12 10000000 230 ns/op 23 B/op 0 allocs/op
BenchmarkCactus-12 3000000 604 ns/op 5 B/op 0 allocs/op
BenchmarkG2s-12 200000 7499 ns/op 576 B/op 21 allocs/op
BenchmarkQuipo-12 1000000 1048 ns/op 384 B/op 7 allocs/op
BenchmarkUnix4ever-12 1000000 1695 ns/op 408 B/op 18 allocs/op
Ideas were borrowed from the following stastd clients:
- https://github.com/quipo/statsd (MIT License, https://github.com/quipo/statsd/blob/master/LICENSE)
- https://github.com/Unix4ever/statsd (MIT License, https://github.com/Unix4ever/statsd/blob/master/LICENSE)
- https://github.com/alexcesaro/statsd/ (MIT License, https://github.com/alexcesaro/statsd/blob/master/LICENSE)
- https://github.com/armon/go-metrics (MIT License, https://github.com/armon/go-metrics/blob/master/LICENSE)
I gave a talk about design and optimizations which went into go-statsd at Gophercon Russia 2018: slides, source.
License is MIT License.