Migrating from 0.31 to 0.32

"Data loaders" renamed to "importers"

The file import system previously called "data loaders" has been renamed to "importers" to avoid confusion with the widely-used ML/PyTorch "dataloader" concept and to better describe what the system does: importing external file formats into Rerun.

The old names are deprecated but most of them still work for this release.

Rust API

All old names are available as deprecated type aliases and will be removed in a future release.

The Cargo feature flag data_loaders has been renamed to importers.

C/C++ API

The old names are still available as deprecated aliases.

Python API

The old name is still available but deprecated.

External importers

Executables on $PATH with the rerun-importer- prefix are now the canonical way to register external importers. The old rerun-loader- prefix continues to work but will log a deprecation warning.

Before:

rerun-loader-my-format

After:

rerun-importer-my-format

Lenses API (Rust)

The Lenses API has been restructured and simplified.

Entity path filtering moved to Lenses

Entity path filtering is no longer part of the Lens itself. Use Lenses::add_lens_with_filter:

let lenses = Lenses::new(OutputMode::DropUnmatched)
    .add_lens_with_filter(EntityPathFilter::parse_forgiving("sensors/**"), lens);

This makes applying lenses to individual chunks more ergonomic.

New builder API

Lenses are now created through Lens::derive(), Lens::scatter(), and Lens::mutate():

// Before
Lens::for_input_column(EntityPathFilter::all(), "component")
    .output_columns(|out| { /* … */ })?
    .build()

// After - derive lens (1:1 row mapping)
Lens::derive("component")
    .to_component(component_descr, ".field")
    .build()?

// After - scatter lens (1:N row mapping)
Lens::scatter("component")
    .output_entity("/target")
    .to_component(component_descr, ".field")
    .build()?

// After - mutate lens (modifies component in-place)
Lens::mutate("component", ".field").build()

To output columns to multiple entities from a single component, multiple lenses can be registered for the same input component.

rerun rrd compact renamed to rerun rrd optimize, has profiles and new defaults

rerun rrd compact is now rerun rrd optimize.

A new --profile argument has been added to opt to known good values. Two profiles are available: live (optimized for the live Viewer workflow, same as previous defaults) and object-store (optimized for querying and streaming from object-store-backed storage, e.g. a catalog server). <!-- NOLINT -->

By default, the object-store profile is now used. Use --profile live to keep the previous behavior. <!-- NOLINT -->

DatasetEntry.register requires a sequence of URIs (Python)

DatasetEntry.register no longer accepts a single URI string for recording_uri. Pass a sequence of URIs instead, and prefer batching many URIs into a single register call rather than calling register repeatedly in a loop (which is much slower).

Old single-string invocations still work at runtime but emit a DeprecationWarning.

- dataset.register(url, layer_name="base")
+ dataset.register([url], layer_name="base")

layer_name is unchanged: pass a single string to apply one layer to all recordings, or a sequence matching the length of recording_uri.

URDF importer transform entity

The URDF importer now loads the static transforms of the model to the /tf_static entity by default. This replaces the model-dependent entity path of previous versions, and improves consistency with ROS data.

A custom entity path can be now also configured in the UrdfTree API in Python and Rust, if desired.

MCAP metadata and statistics

In MCAP to RRD conversion, metadata records, statistics, and recording info are now saved at dedicated reserved entity paths instead of RRD properties (__properties).

Metadata records are saved under __mcap_metadata, and MCAP statistics and recording info are saved under __mcap_properties.

rerun.recording deprecated in favor of RrdReader

The rerun.recording module — Recording, RRDArchive, load_recording, load_archive — is deprecated. Use rerun.experimental.RrdReader instead, which natively supports multi-store RRDs (multiple recordings and blueprints in one file) and lazy loading.