M3

Market position

M3 is presented as an option for cloud native organizations scaling Prometheus-based monitoring, used by large enterprises including Uber, Walmart Labs and LinkedIn, and stated to store tens of billions of active metric time series at the largest production scales.

The project emphasizes horizontal scalability as a single centralized store for monitoring data, drop-in compatibility with Prometheus remote storage and PromQL as well as Graphite and StatsD/Carbon formats, an embedded inverted index for high-cardinality queries, an 11x compression ratio from an optimized compression algorithm, and configurable consistency with three replicas and quorum reads and writes. Cited adopters describe it as cheaper than alternatives at scale and linearly scalable.

Technology

M3DB is written in Go with no required dependencies, optionally using etcd for cluster membership and topology. It was inspired by Gorilla and Cassandra and features distributed time series storage with a WAL commit log, per-shard independent persistence of time windows, an embedded inverted index, synchronous replication with configurable durability and consistency, M3TSZ float64 compression (lossless or lossy), configurable time precision from seconds to nanoseconds, and configurable out-of-order writes bounded by the block size. M3Query is a stateless query engine supporting PromQL and Graphite; M3Coordinator and M3Aggregator handle ingestion and streaming aggregation with per-team resolution and retention policies. Deployment options include Docker, binaries and Kubernetes via the M3DB Operator.

Go-to-market

Adoption is driven through open source distribution on GitHub, Docker images (quay.io/m3db/m3dbnode), a Kubernetes operator, published documentation and quickstart guides, and community channels including Slack, a Google Group forum and community meetings. Public visibility also comes from conference talks and meetups such as Monitorama 2018, FOSDEM 2020 and M3 meetups in June 2020, plus published user case studies.

Cloud native organizations and large enterprises scaling Prometheus-based monitoring, including infrastructure, microservices, machine learning platform and business operations teams; cited users range from startups to companies such as Uber, Walmart Labs and LinkedIn.

Geography

The project is maintained by open source contributors around the globe; cited users include Uber, Walmart Labs (monitoring 5,000 domestic Walmart stores in the United States) and LinkedIn.

History

The project originated in 2014 inside Uber's Observability team, which had been running a legacy metrics platform built on Graphite and WhisperDB. The first iteration of the storage layer used Cassandra and Elasticsearch, but as Uber's ride-hailing business entered hyper growth the operational problems resurfaced, and the team built M3DB — a custom time series database with an embedded inverted index — from scratch over roughly a year and a half, rolling it out across Uber. Slow queries and out-of-memory events then prompted M3Query, a new engine supporting PromQL and the legacy Graphite language. Growing ingestion volumes and divergent retention/resolution requirements across infrastructure and business teams led to M3Coordinator and M3Aggregator for customizable ingestion and streaming aggregation. The project was released as open source by Uber Technologies and reached version 1.0.0; it is now maintained by contributors globally.

Risks & controversies

Documentation notes current limitations that make M3DB unsuitable as a general purpose time series database: the project avoids compactions except in-memory compaction of the mutable compressed time series window (default two hours), so out-of-order writes are limited to a single compressed window and backfilling large amounts of data is not currently possible. Storage and retrieval are optimized for float64 values, so arbitrary data structures are not supported.