feat: curio/lmrpc: Ingest backpressure (filecoin-project#11865)

hanabi1224 · Apr 12, 2024 · 3931710 · 3931710
1 parent 3cc62e0
commit 3931710
Show file tree

Hide file tree

Showing 5 changed files with 209 additions and 27 deletions.
diff --git a/curiosrc/market/lmrpc/lmrpc.go b/curiosrc/market/lmrpc/lmrpc.go
@@ -37,6 +37,8 @@ import (
 
 var log = logging.Logger("lmrpc")
 
+const backpressureWaitTime = 30 * time.Second
+
 func ServeCurioMarketRPCFromConfig(db *harmonydb.DB, full api.FullNode, cfg *config.CurioConfig) error {
 	return forEachMarketRPC(cfg, func(maddr string, listen string) error {
 		addr, err := address.NewFromString(maddr)
@@ -248,46 +250,72 @@ func ServeCurioMarketRPC(db *harmonydb.DB, full api.FullNode, maddr address.Addr
 		var refID int64
 		var pieceWasCreated bool
 
-		comm, err := db.BeginTransaction(ctx, func(tx *harmonydb.Tx) (commit bool, err error) {
-			var pieceID int64
-			// Attempt to select the piece ID first
-			err = tx.QueryRow(`SELECT id FROM parked_pieces WHERE piece_cid = $1`, deal.DealProposal.PieceCID.String()).Scan(&pieceID)
+		for {
+			var backpressureWait bool
 
-			if err != nil {
-				if err == pgx.ErrNoRows {
-					// Piece does not exist, attempt to insert
-					err = tx.QueryRow(`
+			comm, err := db.BeginTransaction(ctx, func(tx *harmonydb.Tx) (commit bool, err error) {
+				// BACKPRESSURE
+				wait, err := maybeApplyBackpressure(tx, conf.Ingest)
+				if err != nil {
+					return false, xerrors.Errorf("backpressure checks: %w", err)
+				}
+				if wait {
+					backpressureWait = true
+					return false, nil
+				}
+
+				var pieceID int64
+				// Attempt to select the piece ID first
+				err = tx.QueryRow(`SELECT id FROM parked_pieces WHERE piece_cid = $1`, deal.DealProposal.PieceCID.String()).Scan(&pieceID)
+
+				if err != nil {
+					if err == pgx.ErrNoRows {
+						// Piece does not exist, attempt to insert
+						err = tx.QueryRow(`
 							INSERT INTO parked_pieces (piece_cid, piece_padded_size, piece_raw_size)
 							VALUES ($1, $2, $3)
 							ON CONFLICT (piece_cid) DO NOTHING
 							RETURNING id`, deal.DealProposal.PieceCID.String(), int64(pieceSize.Padded()), int64(pieceSize)).Scan(&pieceID)
-					if err != nil {
-						return false, xerrors.Errorf("inserting new parked piece and getting id: %w", err)
+						if err != nil {
+							return false, xerrors.Errorf("inserting new parked piece and getting id: %w", err)
+						}
+						pieceWasCreated = true // New piece was created
+					} else {
+						// Some other error occurred during select
+						return false, xerrors.Errorf("checking existing parked piece: %w", err)
 					}
-					pieceWasCreated = true // New piece was created
 				} else {
-					// Some other error occurred during select
-					return false, xerrors.Errorf("checking existing parked piece: %w", err)
+					pieceWasCreated = false // Piece already exists, no new piece was created
 				}
-			} else {
-				pieceWasCreated = false // Piece already exists, no new piece was created
-			}
 
-			// Add parked_piece_ref
-			err = tx.QueryRow(`INSERT INTO parked_piece_refs (piece_id, data_url)
+				// Add parked_piece_ref
+				err = tx.QueryRow(`INSERT INTO parked_piece_refs (piece_id, data_url)
         			VALUES ($1, $2) RETURNING ref_id`, pieceID, dataUrl.String()).Scan(&refID)
+				if err != nil {
+					return false, xerrors.Errorf("inserting parked piece ref: %w", err)
+				}
+
+				// If everything went well, commit the transaction
+				return true, nil // This will commit the transaction
+			}, harmonydb.OptionRetry())
 			if err != nil {
-				return false, xerrors.Errorf("inserting parked piece ref: %w", err)
+				return api.SectorOffset{}, xerrors.Errorf("inserting parked piece: %w", err)
 			}
+			if !comm {
+				if backpressureWait {
+					// Backpressure was applied, wait and try again
+					select {
+					case <-time.After(backpressureWaitTime):
+					case <-ctx.Done():
+						return api.SectorOffset{}, xerrors.Errorf("context done while waiting for backpressure: %w", ctx.Err())
+					}
+					continue
+				}
 
-			// If everything went well, commit the transaction
-			return true, nil // This will commit the transaction
-		}, harmonydb.OptionRetry())
-		if err != nil {
-			return api.SectorOffset{}, xerrors.Errorf("inserting parked piece: %w", err)
-		}
-		if !comm {
-			return api.SectorOffset{}, xerrors.Errorf("piece tx didn't commit")
+				return api.SectorOffset{}, xerrors.Errorf("piece tx didn't commit")
+			}
+
+			break
 		}
 
 		// wait for piece to be parked
@@ -420,3 +448,55 @@ func ServeCurioMarketRPC(db *harmonydb.DB, full api.FullNode, maddr address.Addr
 
 	return server.ListenAndServe()
 }
+
+func maybeApplyBackpressure(tx *harmonydb.Tx, cfg config.CurioIngestConfig) (wait bool, err error) {
+	var bufferedSDR, bufferedTrees, bufferedPoRep int
+	err = tx.QueryRow(`WITH BufferedSDR AS (
+    SELECT SUM(buffered_count) AS buffered_sdr_count
+    FROM (
+             SELECT COUNT(p.task_id_sdr) - COUNT(t.owner_id) AS buffered_count
+             FROM sectors_sdr_pipeline p
+                      LEFT JOIN harmony_task t ON p.task_id_sdr = t.id
+             WHERE p.after_sdr = false
+             UNION ALL
+             SELECT COUNT(1) AS buffered_count
+             FROM parked_pieces
+             WHERE complete = false
+         ) AS subquery
+),
+     BufferedTrees AS (
+         SELECT COUNT(p.task_id_tree_r) - COUNT(t.owner_id) AS buffered_trees_count
+         FROM sectors_sdr_pipeline p
+                  LEFT JOIN harmony_task t ON p.task_id_tree_r = t.id
+         WHERE p.after_sdr = true AND p.after_tree_r = false
+     ),
+     BufferedPoRep AS (
+         SELECT COUNT(p.task_id_porep) - COUNT(t.owner_id) AS buffered_porep_count
+         FROM sectors_sdr_pipeline p
+                  LEFT JOIN harmony_task t ON p.task_id_porep = t.id
+         WHERE p.after_tree_r = true AND p.after_porep = false
+     )
+SELECT
+    (SELECT buffered_sdr_count FROM BufferedSDR) AS total_buffered,
+    (SELECT buffered_trees_count FROM BufferedTrees) AS buffered_trees_count,
+    (SELECT buffered_porep_count FROM BufferedPoRep) AS buffered_porep_count
+`).Scan(&bufferedSDR, &bufferedTrees, &bufferedPoRep)
+	if err != nil {
+		return false, xerrors.Errorf("counting parked pieces: %w", err)
+	}
+
+	if cfg.MaxQueueSDR != 0 && bufferedSDR > cfg.MaxQueueSDR {
+		log.Debugw("backpressure", "reason", "too many SDR tasks", "buffered", bufferedSDR, "max", cfg.MaxQueueSDR)
+		return true, nil
+	}
+	if cfg.MaxQueueTrees != 0 && bufferedTrees > cfg.MaxQueueTrees {
+		log.Debugw("backpressure", "reason", "too many tree tasks", "buffered", bufferedTrees, "max", cfg.MaxQueueTrees)
+		return true, nil
+	}
+	if cfg.MaxQueuePoRep != 0 && bufferedPoRep > cfg.MaxQueuePoRep {
+		log.Debugw("backpressure", "reason", "too many PoRep tasks", "buffered", bufferedPoRep, "max", cfg.MaxQueuePoRep)
+		return true, nil
+	}
+
+	return false, nil
+}
diff --git a/documentation/en/default-curio-config.toml b/documentation/en/default-curio-config.toml
@@ -318,6 +318,37 @@
   #SingleRecoveringPartitionPerPostMessage = false
 
 
+[Ingest]
+  # Maximum number of sectors that can be queued waiting for SDR to start processing.
+  # 0 = unlimited
+  # Note: This mechanism will delay taking deal data from markets, providing backpressure to the market subsystem.
+  # The SDR queue includes deals which are in the process of entering the sealing pipeline - size of this queue
+  # will also impact the maximum number of ParkPiece tasks which can run concurrently.
+  # 
+  # SDR queue is the first queue in the sealing pipeline, meaning that it should be used as the primary backpressure mechanism.
+  #
+  # type: int
+  #MaxQueueSDR = 8
+
+  # Maximum number of sectors that can be queued waiting for SDRTrees to start processing.
+  # 0 = unlimited
+  # Note: This mechanism will delay taking deal data from markets, providing backpressure to the market subsystem.
+  # In case of the trees tasks it is possible that this queue grows more than this limit, the backpressure is only
+  # applied to sectors entering the pipeline.
+  #
+  # type: int
+  #MaxQueueTrees = 0
+
+  # Maximum number of sectors that can be queued waiting for PoRep to start processing.
+  # 0 = unlimited
+  # Note: This mechanism will delay taking deal data from markets, providing backpressure to the market subsystem.
+  # Like with the trees tasks, it is possible that this queue grows more than this limit, the backpressure is only
+  # applied to sectors entering the pipeline.
+  #
+  # type: int
+  #MaxQueuePoRep = 0
+
+
 [Journal]
   # Events of the form: "system1:event1,system1:event2[,...]"
   #

diff --git a/node/config/def.go b/node/config/def.go
@@ -362,5 +362,10 @@ func DefaultCurioConfig() *CurioConfig {
 			PartitionCheckTimeout: Duration(20 * time.Minute),
 			SingleCheckTimeout:    Duration(10 * time.Minute),
 		},
+		Ingest: CurioIngestConfig{
+			MaxQueueSDR:   8, // default to 8 sectors before sdr
+			MaxQueueTrees: 0, // default don't use this limit
+			MaxQueuePoRep: 0, // default don't use this limit
+		},
 	}
 }
diff --git a/node/config/doc_gen.go b/node/config/doc_gen.go
diff --git a/node/config/types.go b/node/config/types.go
@@ -74,6 +74,7 @@ type CurioConfig struct {
 	// Addresses of wallets per MinerAddress (one of the fields).
 	Addresses []CurioAddresses
 	Proving   CurioProvingConfig
+	Ingest    CurioIngestConfig
 	Journal   JournalConfig
 	Apis      ApisConfig
 }
@@ -826,6 +827,31 @@ type CurioProvingConfig struct {
 	SingleRecoveringPartitionPerPostMessage bool
 }
 
+type CurioIngestConfig struct {
+	// Maximum number of sectors that can be queued waiting for SDR to start processing.
+	// 0 = unlimited
+	// Note: This mechanism will delay taking deal data from markets, providing backpressure to the market subsystem.
+	// The SDR queue includes deals which are in the process of entering the sealing pipeline - size of this queue
+	// will also impact the maximum number of ParkPiece tasks which can run concurrently.
+	//
+	// SDR queue is the first queue in the sealing pipeline, meaning that it should be used as the primary backpressure mechanism.
+	MaxQueueSDR int
+
+	// Maximum number of sectors that can be queued waiting for SDRTrees to start processing.
+	// 0 = unlimited
+	// Note: This mechanism will delay taking deal data from markets, providing backpressure to the market subsystem.
+	// In case of the trees tasks it is possible that this queue grows more than this limit, the backpressure is only
+	// applied to sectors entering the pipeline.
+	MaxQueueTrees int
+
+	// Maximum number of sectors that can be queued waiting for PoRep to start processing.
+	// 0 = unlimited
+	// Note: This mechanism will delay taking deal data from markets, providing backpressure to the market subsystem.
+	// Like with the trees tasks, it is possible that this queue grows more than this limit, the backpressure is only
+	// applied to sectors entering the pipeline.
+	MaxQueuePoRep int
+}
+
 // API contains configs for API endpoint
 type API struct {
 	// Binding address for the Lotus API