Uber Redesigned M3DB Shard Placement Model

The company introduced fixed-size subclusters to improve database reliability and limit node failure impacts.

Updated on Sept. 21, 2026 in Data Centers

Isometric editorial illustration of stacked, uniform server chassis modules organized into isolated geometric blocks, representing database infrastructure architecture. — Uber has transitioned its M3DB distributed time series database to a fixed-size subcluster model to improve reliability and isolate node failures. AI Illustration. Upload story photo >

Live Poll

Do you believe complex infrastructure updates improve system reliability for the average user?

Uber has overhauled its M3DB distributed time series database by adopting a fixed-size subcluster placement model. This change replaces previous systems that allowed node failures to disrupt large portions of a cluster.

Why it matters

Previous shard placement models created complex dependency graphs that became difficult to manage as clusters expanded. The new architecture limits the fallout from maintenance or failures to specific, nonoverlapping subcluster groups.

The new design requires subcluster sizes to be multiples of the replication factor, using equal instance weights. In a 12-node cluster with a replication factor of three, the model partitions nodes into fixed-size subclusters of six.

The players

Uber

Uber is a global technology company known for its ride-hailing services and extensive distributed data infrastructure.

The details

The updated system enforces strict isolation groups, ensuring that replicas never share the same rack or availability zone. When scaling, a greedy algorithm moves shards to maintain balanced loads, with destination nodes streaming data from peers before taking ownership.

September 21, 2026: The M3DB redesign details were formally published.

The Tech Race

This redesign reflects an ongoing industry shift toward more modular, failure-resistant database architectures in massive distributed systems. It replaces monolithic scaling approaches that have historically struggled with complex dependency graphs as data footprints grow.

For users and engineers, this change ensures higher database reliability and more predictable performance during routine cluster maintenance. The implementation retains existing placement operations to ensure compatibility with current developer tooling.

The takeaway

Architectural shifts toward smaller, isolated clusters are essential for maintaining uptime in massive distributed environments. Engineers should prioritize shard placement strategies that decouple failure zones to prevent cascading outages.