systemLog:   verbosity: 0   destination: file   path: /data3/yyz_test/mongodb_pipeline/replica_set/37003/logs/mongod.log   logAppend: true   logRotate: renameprocessManagement:   fork: true   pidFilePath: /data3/yyz_test/mongodb_pipeline/replica_set/37003/mongod.pidnet:   port: 37003   bindIpAll: true   maxIncomingConnections: 10000   ipv6: true #  compression: {}   compression:     compressors: zstdsetParameter:   disabledSecureAllocatorDomains: '*'   tcmallocAggressiveMemoryDecommit: 1   cursorTimeoutMillis: 600000   maxTransactionLockRequestTimeoutMillis: 5   internalQueryMaxBlockingSortMemoryUsageBytes: 33554432storage:   dbPath: /data4/yyz_test/mongodb_pipeline/replica_set/37003/data   directoryPerDB: true   syncPeriodSecs: 60   wiredTiger:     engineConfig:       cacheSizeGB: 24       journalCompressor: none     evictionConfig: {}     collectionConfig:       blockCompressor: snappy        replication:   oplogSizeMB: 25600   replSetName: rs0operationProfiling:   mode: "off"   slowOpThresholdMs: 100security:   keyFile: /data3/yyz_test/mongodb_pipeline/replica_set/keyfile   authorization: enabled root@:/data3/yyz_test/mongodb_pipeline# cat /data3/yyz_test/mongodb_pipeline/replica_set/37001/mongod.conf systemLog:   verbosity: 0   destination: file   path: /data3/yyz_test/mongodb_pipeline/replica_set/37001/logs/mongod.log   logAppend: true   logRotate: renameprocessManagement:   fork: true   pidFilePath: /data3/yyz_test/mongodb_pipeline/replica_set/37001/mongod.pidnet:   port: 37001   bindIpAll: true   maxIncomingConnections: 10000   ipv6: true #  compression: {}   compression:     compressors: zstdsetParameter:   disabledSecureAllocatorDomains: '*'   tcmallocAggressiveMemoryDecommit: 1   cursorTimeoutMillis: 600000   maxTransactionLockRequestTimeoutMillis: 5   internalQueryMaxBlockingSortMemoryUsageBytes: 33554432   internalReplPipelineCompressionThreadPoolSize: 10storage:   dbPath: /data2/yyz_test/mongodb_pipeline/replica_set/37001/data   directoryPerDB: true   syncPeriodSecs: 60   wiredTiger:     engineConfig:       cacheSizeGB: 24       journalCompressor: none     evictionConfig: {}     collectionConfig:       blockCompressor: snappy        replication:   oplogSizeMB: 25600   replSetName: rs0operationProfiling:   mode: "off"   slowOpThresholdMs: 100security:   keyFile: /data3/yyz_test/mongodb_pipeline/replica_set/keyfile   authorization: enabled 1 primary + 2 secondary: ./bin/ycsb.sh load mongodb -s -P workloads/workloada -p mongodb.url= 'mongodb: //root:xxxxx@xxxxx:37001,xxxxx:37002,xxxxx:37003/admin?maxPoolSize=500&minPoolSize=500&waitQueueTimeoutMS=120000' -p fieldcount=5000 -threads 100 -p recordcount=100000 -p table=test21 -threads 1000   On a replica set with network (wire) compression enabled, replication sync-source responses ( getMore  on  local.oplog.rs ) are compressed  inline on the connection / session-workflow thread . For workloads dominated by large documents this makes a single replication stream  CPU-bound on one core : the connection thread serializes  read oplog → build response → compress → send , and compression dominates. As a result, replica-set write throughput is capped well below what the storage engine and disk can sustain, even on very fast NVMe. This ticket proposes an  asynchronous, order-preserving compression pipeline : response compression is offloaded from the connection thread to a small shared worker pool, while the connection thread stays the sole sender and preserves strict wire ordering. This lets the sync-source thread keep producing exhaust  getMore  responses while previous batches are compressed in parallel. Measured with a YCSB large-document write load on an 8c/24G node (snappy) backed by ~1M-IOPS NVMe, the pipeline lifts insert throughput from  ~990 ops/s to ~1377 ops/s (about +39%)  and removes the single-connection compression ceiling. With the more CPU-intensive zstd compressor on a 12c/32G node the gain is even larger (~3.85×, see the zstd section).