fix: DefaultSaturnCarRequestTimeout

19s is not enough for fetching CAR stream of unknown length,
every bigger request was failing.

If we need to pick some ceiling, 30m sound like a good starting point
(this is when CAR stream got timeouted on the old ipfs.io).

Author: lidel

caboose.go

@@ -75,11 +75,19 @@ type Config struct {
 	MaxNCoolOff int
 }
 
+const DefaultLoggingInterval = 5 * time.Second
+const DefaultSaturnLoggerRequestTimeout = 1 * time.Minute
+
+const DefaultSaturnOrchestratorRequestTimeout = 19 * time.Second
+
+const DefaultSaturnBlockRequestTimeout = 19 * time.Second
+const DefaultSaturnCarRequestTimeout = 30 * time.Minute
+
 const DefaultMaxRetries = 3
 const DefaultPoolFailureDownvoteDebounce = 1 * time.Minute
 const DefaultPoolMembershipDebounce = 3 * DefaultPoolRefreshInterval
 const DefaultPoolLowWatermark = 5
-const DefaultSaturnRequestTimeout = 19 * time.Second
+
 const maxBlockSize = 4194305 // 4 Mib + 1 byte
 const DefaultOrchestratorEndpoint = "https://orchestrator.strn.pl/nodes/nearby?count=1000"
 const DefaultPoolRefreshInterval = 5 * time.Minute
@@ -188,7 +196,7 @@ func NewCaboose(config *Config) (*Caboose, error) {
 
 	if c.config.SaturnClient == nil {
 		c.config.SaturnClient = &http.Client{
-			Timeout: DefaultSaturnRequestTimeout,
+			Timeout: DefaultSaturnCarRequestTimeout,
 		}
 	}
 	if c.config.OrchestratorEndpoint == nil {

cmd/caboose/main.go

@@ -54,7 +54,7 @@ func main1() int {
 
 				LoggingEndpoint: *le,
 				LoggingClient:   http.DefaultClient,
-				LoggingInterval: 5 * time.Second,
+				LoggingInterval: DefaultLoggingInterval,
 
 				DoValidation: true,
 				PoolRefresh:  5 * time.Minute,

fetcher.go

@@ -81,6 +81,11 @@ func (p *pool) fetchResource(ctx context.Context, from string, resource string,
 	isCacheHit := false
 	networkError := ""
 
+	isBlockRequest := false
+	if mime == "application/vnd.ipld.raw" {
+		isBlockRequest = true
+	}
+
 	defer func() {
 		var ttfbMs int64
 		durationSecs := time.Since(start).Seconds()
@@ -92,15 +97,15 @@ func (p *pool) fetchResource(ctx context.Context, from string, resource string,
 			ttfbMs = fb.Sub(start).Milliseconds()
 			fetchTTFBPerBlockPerPeerSuccessMetric.Observe(float64(ttfbMs))
 			// track individual block metrics separately
-			if mime == "application/vnd.ipld.raw" {
+			if isBlockRequest {
 				fetchDurationPerBlockPerPeerSuccessMetric.Observe(float64(response_success_end.Sub(start).Milliseconds()))
 			} else {
 				fetchDurationPerCarPerPeerSuccessMetric.Observe(float64(response_success_end.Sub(start).Milliseconds()))
 			}
 			fetchSpeedPerBlockPerPeerMetric.Observe(float64(received) / float64(durationMs))
 		} else {
 			fetchTTFBPerBlockPerPeerFailureMetric.Observe(float64(ttfbMs))
-			if mime == "application/vnd.ipld.raw" {
+			if isBlockRequest {
 				fetchDurationPerBlockPerPeerFailureMetric.Observe(float64(time.Since(start).Milliseconds()))
 			} else {
 				fetchDurationPerCarPerPeerFailureMetric.Observe(float64(time.Since(start).Milliseconds()))
@@ -145,7 +150,16 @@ func (p *pool) fetchResource(ctx context.Context, from string, resource string,
 		}
 	}()
 
-	reqCtx, cancel := context.WithTimeout(ctx, DefaultSaturnRequestTimeout)
+	// TODO: Ideally, we would have additional "PerRequestInactivityTimeout"
+	// which is the amount of time without any NEW data from the server, but
+	// that can be added later. We need both because a slow trickle of data
+	// could take a large amount of time.
+	requestTimeout := DefaultSaturnCarRequestTimeout
+	if isBlockRequest {
+		requestTimeout = DefaultSaturnBlockRequestTimeout
+	}
+
+	reqCtx, cancel := context.WithTimeout(ctx, requestTimeout)
 	defer cancel()
 	req, err := http.NewRequestWithContext(reqCtx, http.MethodGet, reqUrl, nil)
 	if err != nil {