Update worker constants for pruning (#1278)

With this update Arequipa has been pruning all night without producing
any "EOF" alerts.
That's because the responses become small enough so that the host
doesn't close the connection.

1000 roots per batch is significantly smaller than the 500k we had
before but the roundabout with `hostd` should be fast enough to not
cause too much of a slowdown. It's also a lot better for `hostd`'s db
since it won't have to commit as many changes at once.

autopilot/contract_pruning.go

@@ -191,7 +191,7 @@ func (ap *Autopilot) pruneContract(w Worker, fcid types.FileContractID, hk types
 	ap.mu.Lock()
 	defer ap.mu.Unlock()
 	alertID := alerts.IDForContract(alertPruningID, fcid)
-	if shouldSendPruneAlert(err, hostVersion) {
+	if shouldSendPruneAlert(err, hostVersion, hostRelease) {
 		ap.RegisterAlert(ctx, newContractPruningFailedAlert(hk, hostVersion, hostRelease, fcid, err))
 		ap.pruningAlertIDs[fcid] = alertID // store id to dismiss stale alerts
 	} else {
@@ -236,8 +236,10 @@ func humanReadableSize(b int) string {
 		float64(b)/float64(div), "KMGTPE"[exp])
 }
 
-func shouldSendPruneAlert(err error, version string) bool {
-	return err != nil && !((utils.IsErr(err, errInvalidMerkleProof) && build.VersionCmp(version, "1.6.0") < 0) ||
+func shouldSendPruneAlert(err error, version, release string) bool {
+	merkleRootIssue := utils.IsErr(err, errInvalidMerkleProof) &&
+		(build.VersionCmp(version, "1.6.0") < 0 || version == "1.6.0" && release == "")
+	return err != nil && !(merkleRootIssue ||
 		utils.IsErr(err, utils.ErrConnectionRefused) ||
 		utils.IsErr(err, utils.ErrConnectionTimedOut) ||
 		utils.IsErr(err, utils.ErrConnectionResetByPeer) ||

autopilot/contractor/alerts.go

@@ -41,14 +41,16 @@ func newContractRenewalFailedAlert(contract api.ContractMetadata, interrupted bo
 	}
 }
 
-func newLostSectorsAlert(hk types.PublicKey, lostSectors uint64) alerts.Alert {
+func newLostSectorsAlert(hk types.PublicKey, version, release string, lostSectors uint64) alerts.Alert {
 	return alerts.Alert{
 		ID:       alerts.IDForHost(alertLostSectorsID, hk),
 		Severity: alerts.SeverityWarning,
 		Message:  "Host has lost sectors",
 		Data: map[string]interface{}{
 			"lostSectors": lostSectors,
 			"hostKey":     hk.String(),
+			"hostVersion": version,
+			"hostRelease": release,
 			"hint":        "The host has reported that it can't serve at least one sector. Consider blocking this host through the blocklist feature. If you think this was a mistake and you want to ignore this warning for now you can reset the lost sector count",
 		},
 		Timestamp: time.Now(),

autopilot/contractor/contractor.go

@@ -299,7 +299,7 @@ func (c *Contractor) performContractMaintenance(ctx *mCtx, w Worker) (bool, erro
 	var toDismiss []types.Hash256
 	for _, h := range hosts {
 		if registerLostSectorsAlert(h.Interactions.LostSectors*rhpv2.SectorSize, h.StoredData) {
-			c.alerter.RegisterAlert(ctx, newLostSectorsAlert(h.PublicKey, h.Interactions.LostSectors))
+			c.alerter.RegisterAlert(ctx, newLostSectorsAlert(h.PublicKey, h.Settings.Version, h.Settings.Release, h.Interactions.LostSectors))
 		} else {
 			toDismiss = append(toDismiss, alerts.IDForHost(alertLostSectorsID, h.PublicKey))
 		}

worker/rhpv2.go

@@ -25,9 +25,9 @@ const (
 	minMessageSize = 4096
 
 	// maxMerkleProofResponseSize caps the response message size to a generous
-	// 32 MiB max length since batchSizeDeleteSectors assumes ~16MiB of
-	// roots. So we double that to be safe.
-	maxMerkleProofResponseSize = 8 * 1 << 22 // 32 MiB
+	// value of 100 MB worth of roots. This is approximately double the size of
+	// what we have observed on the live network for 5TB+ contracts to be safe.
+	maxMerkleProofResponseSize = 100 * 1 << 20 // 100 MB
 )
 
 var (

worker/worker.go

@@ -35,8 +35,8 @@ import (
 )
 
 const (
-	batchSizeDeleteSectors = uint64(500000) // ~16MiB of roots
-	batchSizeFetchSectors  = uint64(130000) // ~4MiB of roots
+	batchSizeDeleteSectors = uint64(1000)  // 4GiB of contract data
+	batchSizeFetchSectors  = uint64(25600) // 100GiB of contract data
 
 	defaultLockTimeout          = time.Minute
 	defaultRevisionFetchTimeout = 30 * time.Second