Self-aware Golang profile dumper.
Our online system often crashes at midnight (usually killed by the OS due to OOM). As lazy developers, we don't want to be woken up at midnight and waiting for the online error to recur.
holmes comes to rescue.
Holmes collects the following stats every interval passed:
runtime.NumGoroutine
.In addition, holmes will collect RSS
based on GC cycle, if you enable GC heap
.
After warming up(10 times collects after starting application) phase finished, Holmes will compare the current stats with the average of previous collected stats(10 cycles). If the dump rule is matched, Holmes will dump the related profile to log(text mode) or binary file(binary mode).
When you get warning messages sent by your own monitor system, e.g, memory usage exceed 80%, OOM killed, CPU usage exceed 80%, goroutine num exceed 100k. The profile is already dumped to your dump path. You could just fetch the profile and see what actually happened without pressure.
go get mosn.io/holmes
h, _ := holmes.New(
holmes.WithCollectInterval("5s"),
holmes.WithDumpPath("/tmp"),
holmes.WithTextDump(),
holmes.WithDumpToLogger(true),
holmes.WithGoroutineDump(10, 25, 2000, 10*1000,time.Minute),
)
h.EnableGoroutineDump()
// start the metrics collect and dump loop
h.Start()
// stop the dumper
h.Stop()
/tmp
dir.&& current_goroutine_num >
125%* previous_average_goroutine_num or current_goroutine_num >
2000,
time.Minute` means once a dump happened, the next dump will not happen before cooldown
finish-1 minute.
WithGoroutineDump(min int, diff int, abs int, max int, coolDown time.Duration) 100*1000 means max goroutine number, when current goroutines number is greater 100k, holmes would not dump goroutine profile. Cuz if goroutine num is huge, e.g, 100k goroutine dump will also become a heavy action: stw && stack dump. Max = 0 means no limit.
h, _ := holmes.New(
holmes.WithCollectInterval("5s"),
holmes.WithDumpPath("/tmp"),
holmes.WithCPUDump(20, 25, 80, time.Minute),
holmes.WithCPUMax(90),
)
h.EnableCPUDump()
// start the metrics collect and dump loop
h.Start()
// stop the dumper
h.Stop()
/tmp
dir.10%
&&
cpu usage > 125%
* previous cpu usage recorded or cpu usage > 80%
.
time.Minute
means once a dump happened, the next dump will not happen before
cooldown finish-1 minute.h, _ := holmes.New(
holmes.WithCollectInterval("5s"),
holmes.WithDumpPath("/tmp"),
holmes.WithTextDump(),
holmes.WithMemDump(30, 25, 80,time.Minute),
)
h.EnableMemDump()
// start the metrics collect and dump loop
h.Start()
// stop the dumper
h.Stop()
/tmp
dir.10%
&&
memory usage > 125%
* previous memory usage or memory usage > 80%
.
time.Minute
means once a dump happened, the next dump will not happen before
cooldown finish-1 minute.In some situations we can not get useful information, such the application allocates heap memory and
collects it between one CollectInterval
. So we design a new heap memory monitor rule, which bases on
GC cycle, to control holmes dump. It will dump twice heap profile continuously while RSS spike, then devs
can compare the profiles through pprof base
command.
h, _ := holmes.New(
holmes.WithDumpPath("/tmp"),
holmes.WithLogger(holmes.NewFileLog("/tmp/holmes.log", mlog.INFO)),
holmes.WithBinaryDump(),
holmes.WithMemoryLimit(100*1024*1024), // 100MB
holmes.WithGCHeapDump(10, 20, 40, time.Minute),
// holmes.WithProfileReporter(reporter),
)
h.EnableGCHeapDump().Start()
time.Sleep(time.Hour)
You can use Set
method to modify holmes' configurations when the application is running.
h.Set(
WithCollectInterval("2s"),
WithGoroutineDump(min, diff, abs, 90, time.Minute))
You can use Reporter
to implement the following features:
type ReporterImpl struct{}
func (r *ReporterImpl) Report(pType string, filename string, reason ReasonType, eventID string, sampleTime time.Time, pprofBytes []byte, scene Scene) error{
// do something
}
......
r := &ReporterImpl{} // a implement of holmes.ProfileReporter Interface.
h, _ := holmes.New(
holmes.WithProfileReporter(reporter),
holmes.WithDumpPath("/tmp"),
holmes.WithLogger(holmes.NewFileLog("/tmp/holmes.log", mlog.INFO)),
holmes.WithBinaryDump(),
holmes.WithMemoryLimit(100*1024*1024), // 100MB
holmes.WithGCHeapDump(10, 20, 40, time.Minute),
)
Holmes supports to upload your profile to pyroscope server. More details click here please.
Noted that NOT set TextDump
when you enable holmes as pyroscope client.
It's easy.
h, _ := holmes.New(
holmes.WithCollectInterval("5s"),
holmes.WithDumpPath("/tmp"),
holmes.WithTextDump(),
holmes.WithCPUDump(10, 25, 80, time.Minute),
holmes.WithMemDump(30, 25, 80, time.Minute),
holmes.WithGCHeapDump(10, 20, 40, time.Minute),
holmes.WithGoroutineDump(500, 25, 20000, 0, time.Minute),
)
h.EnableCPUDump().
EnableGoroutineDump().
EnableMemDump().
EnableGCHeapDump().Start()
h, _ := holmes.New(
holmes.WithCollectInterval("5s"),
holmes.WithDumpPath("/tmp"),
holmes.WithTextDump(),
holmes.WithCPUDump(10, 25, 80,time.Minute),
holmes.WithCGroup(true), // set cgroup to true
)
If golang version < 1.19, collect a goroutine itself may cause latency spike because of the long time STW. At golang 1.19, it has been optz by concurrent way at this CL.
See our contributor guide.
Scan the QR code below with DingTalk(钉钉) to join the Holmes user group.