coder · spikecurtis · Feb 15, 2024 · Feb 6, 2024 · Feb 9, 2024 · Feb 7, 2024
diff --git a/agent/logs.go b/agent/logs.go
@@ -0,0 +1,239 @@
+package agent
+
+import (
+	"context"
+	"sync"
+	"time"
+
+	"github.com/google/uuid"
+	"golang.org/x/xerrors"
+
+	"cdr.dev/slog"
+	"github.com/coder/coder/v2/agent/proto"
+	"github.com/coder/coder/v2/codersdk/agentsdk"
+)
+
+const (
+	flushInterval    = time.Second
+	maxBytesPerBatch = 1 << 20 // 1MiB
+	overheadPerLog   = 21      // found by testing
+
+	// maxBytesQueued is the maximum length of logs we will queue in memory.  The number is taken
+	// from dump.sql `max_logs_length` constraint, as there is no point queuing more logs than we'll
+	// accept in the database.
+	maxBytesQueued = 1048576
+)
+
+type logQueue struct {
+	logs           []*proto.Log
+	flushRequested bool
+	lastFlush      time.Time
+}
+
+// logSender is a subcomponent of agent that handles enqueuing logs and then sending them over the
+// agent API.  Things that need to log call enqueue and flush.  When the agent API becomes available,
+// the agent calls sendLoop to send pending logs.
+type logSender struct {
+	*sync.Cond
+	queues           map[uuid.UUID]*logQueue
+	logger           slog.Logger
+	exceededLogLimit bool
+	outputLen        int
+}
+
+type logDest interface {
+	BatchCreateLogs(ctx context.Context, request *proto.BatchCreateLogsRequest) (*proto.BatchCreateLogsResponse, error)
+}
+
+func newLogSender(logger slog.Logger) *logSender {
+	return &logSender{
+		Cond:   sync.NewCond(&sync.Mutex{}),
+		logger: logger,
+		queues: make(map[uuid.UUID]*logQueue),
+	}
+}
+
+func (l *logSender) enqueue(src uuid.UUID, logs ...agentsdk.Log) {
+	logger := l.logger.With(slog.F("log_source_id", src))
+	if len(logs) == 0 {
+		logger.Debug(context.Background(), "enqueue called with no logs")
+		return
+	}
+	l.L.Lock()
+	defer l.L.Unlock()
+	if l.exceededLogLimit {
+		logger.Warn(context.Background(), "dropping enqueued logs because we have reached the server limit")
+		// don't error, as we also write to file and don't want the overall write to fail
+		return
+	}
+	defer l.Broadcast()
+	q, ok := l.queues[src]
+	if !ok {
+		q = &logQueue{}
+		l.queues[src] = q
+	}
+	for k, log := range logs {
+		// Here we check the queue size before adding a log because we want to queue up slightly
+		// more logs than the database would store to ensure we trigger "logs truncated" at the
+		// database layer.  Otherwise, the end user wouldn't know logs are truncated unless they
+		// examined the Coder agent logs.
+		if l.outputLen > maxBytesQueued {
+			logger.Warn(context.Background(), "log queue full; truncating new logs", slog.F("new_logs", k), slog.F("queued_logs", len(q.logs)))
+			return
+		}
+		pl, err := agentsdk.ProtoFromLog(log)
+		if err != nil {
+			logger.Critical(context.Background(), "failed to convert log", slog.Error(err))
+			return
+		}
+		if len(pl.Output)+overheadPerLog > maxBytesPerBatch {
+			logger.Warn(context.Background(), "dropping log line that exceeds our limit", slog.F("len", len(pl.Output)))
+			continue
+		}
+		q.logs = append(q.logs, pl)
+		l.outputLen += len(pl.Output)
+	}
+	logger.Debug(context.Background(), "enqueued agent logs", slog.F("new_logs", len(logs)), slog.F("queued_logs", len(q.logs)))
+}
+
+func (l *logSender) flush(src uuid.UUID) {
+	l.L.Lock()
+	defer l.L.Unlock()
+	defer l.Broadcast()
+	q, ok := l.queues[src]
+	if ok {
+		q.flushRequested = true
+	}
+	// queue might not exist because it's already been flushed and removed from
+	// the map.
+}
+
+// sendLoop sends any pending logs until it hits an error or the context is canceled.  It does not
+// retry as it is expected that a higher layer retries establishing connection to the agent API and
+// calls sendLoop again.
+func (l *logSender) sendLoop(ctx context.Context, dest logDest) error {
+	l.L.Lock()
+	defer l.L.Unlock()
+	if l.exceededLogLimit {
+		l.logger.Debug(ctx, "aborting sendLoop because log limit is already exceeded")
+		// no point in keeping this loop going, if log limit is exceeded, but don't return an
+		// error because we're already handled it
+		return nil
+	}
+
+	ctxDone := false
+	defer l.logger.Debug(ctx, "sendLoop exiting")
+
+	// wake 4 times per flush interval to check if anything needs to be flushed
+	ctx, cancel := context.WithCancel(ctx)
+	defer cancel()
+	go func() {
+		tkr := time.NewTicker(flushInterval / 4)
+		defer tkr.Stop()
+		for {
+			select {
+			// also monitor the context here, so we notice immediately, rather
+			// than waiting for the next tick or logs
+			case <-ctx.Done():
+				l.L.Lock()
+				ctxDone = true
+				l.L.Unlock()
+				l.Broadcast()
+				return
+			case <-tkr.C:
+				l.Broadcast()
+			}
+		}
+	}()
+
+	for {
+		for !ctxDone && !l.hasPendingWorkLocked() {
+			l.Wait()
+		}
+		if ctxDone {
+			return nil
+		}
+
+		src, q := l.getPendingWorkLocked()
+		logger := l.logger.With(slog.F("log_source_id", src))
+		q.flushRequested = false // clear flag since we're now flushing
+		req := &proto.BatchCreateLogsRequest{
+			LogSourceId: src[:],
+		}
+
+		// outputToSend keeps track of the size of the protobuf message we send, while
+		// outputToRemove keeps track of the size of the output we'll remove from the queues on
+		// success.  They are different because outputToSend also counts protocol message overheads.
+		outputToSend := 0
+		outputToRemove := 0
+		n := 0
+		for n < len(q.logs) {
+			log := q.logs[n]
+			outputToSend += len(log.Output) + overheadPerLog
+			if outputToSend > maxBytesPerBatch {
+				break
+			}
+			req.Logs = append(req.Logs, log)
+			n++
+			outputToRemove += len(log.Output)
+		}
+
+		l.L.Unlock()
+		logger.Debug(ctx, "sending logs to agent API", slog.F("num_logs", len(req.Logs)))
+		resp, err := dest.BatchCreateLogs(ctx, req)
E864		l.L.Lock()
+		if err != nil {
+			return xerrors.Errorf("failed to upload logs: %w", err)
+		}
+		if resp.LogLimitExceeded {
+			l.logger.Warn(ctx, "server log limit exceeded; logs truncated")
+			l.exceededLogLimit = true
+			// no point in keeping anything we have queued around, server will not accept them
+			l.queues = make(map[uuid.UUID]*logQueue)
+			// We've handled the error as best as we can. We don't want the server limit to grind
+			// other things to a halt, so this is all we can do.
+			return nil
+		}
+
+		// Since elsewhere we only append to the logs, here we can remove them
+		// since we successfully sent them.  First we nil the pointers though,
+		// so that they can be gc'd.
+		for i := 0; i < n; i++ {
+			q.logs[i] = nil
+		}
+		q.logs = q.logs[n:]
+		l.outputLen -= outputToRemove
+		if len(q.logs) == 0 {
+			// no empty queues
+			delete(l.queues, src)
+			continue
+		}
+		q.lastFlush = time.Now()
+	}
+}
+
+func (l *logSender) hasPendingWorkLocked() bool {
+	for _, q := range l.queues {
+		if time.Since(q.lastFlush) > flushInterval {
+			return true
+		}
+		if q.flushRequested {
+			return true
+		}
+	}
+	return false
+}
+
+func (l *logSender) getPendingWorkLocked() (src uuid.UUID, q *logQueue) {
+	// take the one it's been the longest since we've flushed, so that we have some sense of
+	// fairness across sources
+	var earliestFlush time.Time
+	for is, iq := range l.queues {
+		if q == nil || iq.lastFlush.Before(earliestFlush) {
+			src = is
+			q = iq
+			earliestFlush = iq.lastFlush
+		}
+	}
+	return src, q
+}