func (l *evtLogger) Log(ctx context.Context, evt *crit_event.ChromeInfraEvent) {
// Get this as soon as possible, so proto encoding time etc
// doesn't bump the logged event time.
clck := clock.Get(ctx)
now := clck.Now().Unix()
evt.EventSource = l.src
// TODO: Consider making this a func parameter.
evt.TimestampKind = crit_event.ChromeInfraEvent_POINT.Enum()
if trace, ok := ctx.Value(TraceID).(string); ok {
evt.TraceId = proto.String(trace)
}
// TODO:
// evt.SpanId = // ?
// evt.ParentId = // ?
bytes, err := proto.Marshal(evt)
if err != nil {
logging.Errorf(ctx, "Error marshaling ChromeInfraEvent: %v", err.Error())
return
}
l.evtCh <- &crit_event.LogRequestLite_LogEventLite{
EventTimeMs: proto.Int64(now),
EventCode: proto.Int32(0),
EventFlowId: proto.Int32(0),
SourceExtension: bytes,
}
}
// Start runs the event flushing loop. It does not return until either
// an error occurs or ctx.Done().
func (l *evtLogger) start(ctx context.Context) error {
defer close(l.evtCh)
evtBuf := []*crit_event.LogRequestLite_LogEventLite{}
clck := clock.Get(ctx)
flush := func(evts []*crit_event.LogRequestLite_LogEventLite) error {
// TODO: Auth, retry/backoff etc.
req := &crit_event.LogRequestLite{
RequestTimeMs: proto.Int64(clck.Now().Unix()),
LogSourceName: proto.String("CHROME_INFRA"),
LogEvent: evts,
}
logging.Infof(ctx, "Flushing %v events\n", len(req.LogEvent))
b, err := proto.Marshal(req)
if err != nil {
return err
}
resp, err := l.cfg.Client.Post(l.cfg.Endpoint, "application/octet-stream", bytes.NewBuffer(b))
if err != nil {
return err
}
defer resp.Body.Close()
if resp.StatusCode >= 400 {
b, err := httputil.DumpResponse(resp, true)
logging.Errorf(ctx, "Error body: %v %v", string(b), err)
return fmt.Errorf("http status %d: %s", resp.StatusCode, l.cfg.Endpoint)
}
return nil
}
// Loop:
// - If you can read an event off ch, append it to the array of events.
// - If you can't read an event off ch, and it's been at least 1s since the
// last flush call, then flush the events if there are any buffered up.
// - if ctx.Done() returns something, try one last flush and exit.
// Ways this can act poorly:
// - If the capacity of l.evtCh is full and a client attempts to log an event,
// that client will block. This is unlikely to happen because the select
// statement spends very little time on anything else, including the flush
// operation.
// - If flush operations take longer than the flushTicker delay, then
// multiple simultaneous flush operations can stack up.
// - If flush operation fail, there is no retry mechanism (yet) and they
// will just log an error locally.
ticker := clck.After(l.cfg.Interval)
for {
select {
case evt := <-l.evtCh:
evtBuf = append(evtBuf, evt)
case <-ticker:
if len(evtBuf) > 0 {
// Flush in a separate goroutine so we can free this
// one up to keep reading off of l.evtCh and not block
// clients. TODO(seanmccullough): add retry logic on
// POST failure, but make sure to not leak memory. This
// may mean dropping events or persisting them to disk.
go func(evts []*crit_event.LogRequestLite_LogEventLite) {
if err := flush(evts); err != nil {
logging.Errorf(ctx, "Error flushing: %v", err.Error())
}
}(evtBuf)
// The above goroutine has a copy of evtBuf, so
// we can now reset it safely.
evtBuf = []*crit_event.LogRequestLite_LogEventLite{}
}
ticker = clck.After(time.Second)
case <-ctx.Done():
if len(evtBuf) > 0 {
if err := flush(evtBuf); err != nil {
logging.Errorf(ctx, "Error flushing: %v", err.Error())
}
}
logging.Infof(ctx, "Event logger exiting.")
return nil
}
}
}