"Earlier designs of Datadog's storage infrastructure separated responsibilities across multiple systems. Metrics data was written into the Real-Time Database (RTDB), which stored <timeseries_id, timestamp, value> tuples, while an index database maintained identifiers and tags. A storage router directed metrics to RTDB nodes, and queries fanned out across RTDB and index nodes. Each node contained subsystems for ingestion, storage, snapshots, queries, and throttling, all coordinated through a shared control plane."
"This architecture of RTDB went through several iterations. The first generation relied on Cassandra for high write throughput but limited query flexibility. A Redis-based design followed, improving responsiveness but encountering durability and single-thread execution issues. MDBM, a memory-mapped key-value store, offered better use of operating system caching but ran into scalability bottlenecks. A subsequent Go-based B+ tree engine adopted a thread-per-core model, which added concurrency but also complexity. Later, RocksDB provided persistence and support for distribution metrics through DDSketch, though challenges in scaling remained."
Datadog launched Monocle, a Rust-based real-time timeseries storage engine that unifies metrics storage, increases ingestion throughput, and lowers query latency while simplifying operations. Earlier designs split responsibilities across RTDB, an index database, a storage router, and per-node subsystems for ingestion, storage, snapshots, queries, and throttling, creating concurrency and scaling challenges. Historical backends included Cassandra, Redis, MDBM, a Go B+ tree, and RocksDB with DDSketch, each with tradeoffs. Monocle uses a shard-per-core, worker-per-shard model with per-worker LSM-trees to ingest data, serve queries, and perform background tasks like compaction.
Read at InfoQ
Unable to calculate read time
Collection
[
|
...
]