Rerun Assess

data-platform observability open-source physical-ai robotics multimodal-data time-series observability-and-visualization columnar-storage ros2 open-source-sdk

Sep 2026

Overview

Rerun is an open-source data layer for physical AI: logging, storage, query, visualization and streaming for multi-rate, multimodal data. Its data model is built around ingesting images, point clouds, transforms, time series, joint states and video from robot logs, human-data rigs, simulation and web video, in formats including MCAP, its own rrd files, and LeRobot datasets. Data is stored as column-chunks in rrd files, and the vendor positions the stack as spanning collection, refinement, training and deployment of robotics data.

We place Rerun in assess because it fills a real gap. Our existing data-platform entries — lakehouse tables, vector stores, feature stores — assume web-shaped or tabular data and struggle with time-, space- and embodiment-tied sensor streams. Rerun's project lead frames exactly this problem: existing infrastructure was largely built around web data, and robot learning needs a unified layer for multimodal, multi-rate streams. With the 0.32 SDK release the project moved beyond its original reputation as a visualization tool toward a unified data layer covering visualization, querying, transformation and training.

Assess, not trial, reflects maturity rather than doubt about the direction. The open-source SDK is easy to adopt for inspection and debugging today; the production backend (Rerun Hub) is a commercial private-preview product, and key ingestion and catalog paths are still marked experimental. Teams should run bounded evaluations and form their own view before betting a data loop on it.

Adoption Signals

The 0.32 SDK release is described as making Rerun a unified data layer for physical data — visualization, querying, transformation and training of multi-rate, multimodal robotics data — a clear widening of scope beyond visualization.
Sustained release cadence in the GitHub releases feed: 0.30 in February 2026, 0.35 in July 2026, 0.36.2 in August 2026 and 0.37.0 in September 2026, with prebuilt CLI, C, C++ and web-viewer artifacts per release.
Native robotics ingestion is being invested in: 0.26 lets you run rerun your-ros2-bag.mcap directly, and the changelog shows 0.35 adding time-windowed and corrupted MCAP conversion, improved ROS 2 timestamp handling, HDF5 import and preliminary LeRobot v3 support.
Three supported languages (C++, Python, Rust) and 10K+ GitHub stars are cited on the product page, alongside an end-to-end collection-to-training positioning.
A commercial offering exists: Rerun Hub, the catalog and backend for large-scale storage, access and streaming from object storage, was announced as available in private preview in May 2026.
Streaming and query ergonomics have matured incrementally — 0.24 added full support for streaming encoded video, and 0.30 added plotting of arbitrary scalars including values from MCAP files and on-demand streaming.

Risks

Experimental surfaces on the critical path. MCAP support is described as still experimental and subject to change, and the built-in Viewer catalog introduced in 0.35 must be enabled via settings and is acknowledged to have rough edges.
Commercial dependency for scale. The open-source SDK is the on-ramp, but Rerun Hub — the catalog, byte-range indexing and retrieval layer that makes the SDK scale over object storage — is a vendor product that was in private preview, so pricing, SLAs and exit options cannot be assessed from public evidence.
Breaking changes between minor versions. Releases ship migration guides and remove or change APIs (0.35 altered the Rust StateChange::with_state signature and dropped the CLI --follow mode for tailing .rrd files), which imposes ongoing upgrade cost on anything embedded in pipelines.
Format and lock-in considerations. The data layer is built on the rrd column-chunk file format; while conversion from other formats is advertised, teams should verify round-trip guarantees before making rrd the system of record.
Name collision. An unrelated product also called Rerun (rerun.build, operated by Reunit SA) markets an autonomous AI agent platform; expect confusion in procurement, search and internal documentation, and reference the vendor explicitly as rerun.io.
Narrow applicability. The value proposition is tightly coupled to physical-AI workloads; teams without sensor, robot or CV data will find little that existing platform entries do not already cover.

Pros & Cons

Advantages

Rerun handles multi-rate, multimodal data such as images, point clouds, transforms, time series, joint states and video as a first-class data model, which the lakehouse, vector-store and feature-store entries on our radar do not address.
Getting started is genuinely cheap: pip install rerun-sdk or cargo add rerun gives you the viewer, query library, CLI and file format, and you can point the CLI straight at a ROS2 bag with rerun your-ros2-bag.mcap.
The project ships frequently and visibly, with a public changelog and release cadence running from 0.30 in February 2026 through 0.37 in September 2026, plus C++, Python and Rust SDKs for mixed-language robotics stacks.

Disadvantages

The scaling story depends on Rerun Hub, the commercial backend, which was announced as private preview in May 2026 and is therefore not something you can evaluate in the open or self-serve.
Several of the ingestion and catalog capabilities teams would rely on are explicitly labelled experimental — MCAP support is 'still experimental and subject to change' and the Viewer catalog has to be enabled in settings and still has rough edges.
Releases carry frequent breaking changes with migration guides (for example 0.35 changed Rust StateChange::with_state and removed the CLI --follow tailing mode), so pinning versions and budgeting for upgrade work is unavoidable.

Recommendation

If you have robotics, teleop, autonomous-vehicle or computer-vision workloads, start with a low-commitment evaluation of the open-source SDK. Install it (pip install rerun-sdk) and point the CLI at an existing ROS2 bag or MCAP file to see how well your real recordings survive ingestion, remembering that MCAP support is still flagged experimental. This alone often pays for itself as a debugging and inspection tool, independent of any decision about the wider data layer.

Only after that should you probe the platform claims. Take one representative dataset through refine-and-query workflows and, if streaming to training is the goal, test the dataframe/query path at realistic volume — this is where the dependency on Rerun Hub becomes concrete, and where you should engage the vendor about private-preview terms, roadmap and self-hosting options. Pin exact SDK versions in CI and read the migration guide for each minor bump; the 0.35 breaking changes are a fair indication of the ongoing cost.

Do not yet make rrd your system of record for data you cannot regenerate, and keep the canonical logs in their original format until conversion fidelity is proven in your own tests. Revisit at the next radar cycle, with promotion to trial contingent on Rerun Hub reaching general availability and MCAP ingestion losing its experimental label. Finally, disambiguate the name in internal docs and vendor records, since an unrelated agent platform ships under the same brand.