Hybrid transactional and analytical workloads

MatrixOne runs transactional and analytical queries on one transactional data system. Separating compute, transactions, logs and object storage helps manage mixed workloads; capacity/isolation still depend on plans, CN groups and shared services. Baseline: v4.2.5.

Data and commit

TN establishes durability for small writes through WAL/LogService, with later object flush. Bulk paths may write objects on CN and commit their metadata. TN supplies committed rows, deletions and object state through Logtail to subscribed CNs. Logtail is not the WAL left after truncation, and CN does not subscribe to Logtail from LogService.

Plans read snapshot-visible objects together with memory changes, deletions and workspace through FileService. A universal algorithm of returning on a Logtail hit and otherwise trying caches/S3 is incorrect. Commit, another CN applying Logtail and object flush occur separately.

Workloads and scaling

Proxy tenant/label groups route sessions; distributed plans place operators; remote cache RPCs fetch file data. Observe these separately. CN groups allocate compute resources while TN, log quorum and object storage still need shared-capacity planning.

Verification

Fix versions, hardware, distribution and plans. Record transaction success/latency, analytical throughput/latency, lock/commit waits, Logtail progress, scan volume, caches and object I/O. Compare separate and simultaneous workloads and define availability/recovery targets during failures. One faster query does not establish linear scaling or isolation.

See architecture, transactions/locks, caches and CN groups.