Skip to content

Releases

These are the releases of the cluster software for clusters you run with the Helm chart or on bare-metal. Each release pins one polars version and is listed with the polars-nanook client release that pins the same version, so that is the client release to install. Client releases up to 0.12 were published as polars-cloud. The client's own release notes are on GitHub. Access the released artifacts lists where to get the chart and the package.

Cluster software releases

0.10.0 (latest) - 5 Nov 2026

⚠ Polars On-Prem is now Polars Nanook, which renames the package, the container image and the environment variables. See Migrate from Polars Cloud and Polars On-Prem.

Highlights

  • Join runtime filters, on by default, drop the rows that can't find a match before they are shuffled
  • Distributed group_by(maintain_order=True), unique(maintain_order=True), joins with maintain_order, and concat, keeping the row order
  • Distributed quantile and median, skew, kurtosis, value_counts, top_k_by, arg_min/arg_max, arg_where, index_of, forward_fill/backward_fill, interpolate, gather_every, row_index(), shift, diff and pct_change with a literal n, and concat(how="horizontal")
  • Distributed sort by keys and group_by keys that aren't elementwise, such as pl.col("a").shift()
  • Group-bys with top_k, top_k_by or arg_where aggregate before the shuffle
  • As-of joins without by can read only the right-side rows each partition can match, instead of broadcasting the right side
  • Better query plans from parquet statistics
  • Int128 and UInt128 columns in collected results
  • Clusters log a warning for each config field they ignore, such as a misspelled key
  • The bare-metal package requires Python 3.11 or later
  • Unstable support for running queries on GPUs on bare-metal clusters

0.9.1 - 15 Oct 2026

Highlights

  • Distributed .over() window expressions:
    • aggregations that return one value per group, such as sum, mean, len, n_unique, first and last
    • any window function, such as cum_sum or shift, when a later sort, group_by or join doesn't need the input row order kept
  • Distributed min_by/max_by, implode, explode, gather, and rolling_corr/rolling_cov
  • Support for out-of-core

0.9.0 - 7 Oct 2026

Highlights

  • Remote explain() to get the query plan without running the query
  • Query priority
  • Distributed product, str.join, is_sorted, approx_quantile, peak_min/peak_max, group_by_dynamic without keys and rolling group-bys with any aggregation
  • Faster shuffles and lower per-stage latency
  • Scans balanced by file size across workers, with large parquet files split by row group
  • Iceberg sinks require pyiceberg >= 0.12 on the workers

0.8.7 - 22 Sept 2026

Highlights

  • Support for collect and collect_batches on ClusterContext
  • Distributed rolling_min, rolling_max, ..., rolling_<agg> expressions
  • Support for worker autoscaling
  • Improved query planning for multistage plans
  • Improved performance and distribution for shuffles
  • Better metrics reporting, and error reporting for resource-starved queries

0.7.1 - 3 Aug 2026

Highlights

  • Stream distributed query results through sink_batches()
  • New experimental Miso query planner, opt-in via distributed(planner="miso")
  • Distributed pl.collect_all()
  • Experimental HDFS support for scans and Iceberg metadata
  • Anonymous query results can be deleted when allow_delete is configured

Older cluster releases

Version Date Polars / client Highlights
0.6.3 1 July 2026 1.42.1 / 0.9.0 Distributed expressions, improved ClusterContext API, distributed Iceberg sink, manual cluster scaling and disk I/O metrics
0.5.2 16 June 2026 1.41.2 / 0.8.0 Externally scaled clusters, faster presigning of anonymous results and cgroup setup improvements
0.5.1 3 June 2026 1.41.2 / 0.8.0 Maintenance release (see the client 0.8.0 notes)
0.4.3 29 May 2026 1.40.1 / 0.7.0 Custom environment variables, shuffle optimizations, and distributed Gather and row-index scans
0.3.1 29 Apr 2026 1.40.1 / 0.6.1 Distributed row index, row-index scan lowering and pre-slice in row count estimation
0.2.4 13 Apr 2026 1.39.3 / 0.6.0 Single-node expression lowering, better detection of hanging queries and distributed slice
0.2.3 23 Mar 2026 1.39.3 / 0.6.0 Scratchpad, per-query historical profiling data in the control plane, and endpoints to download the config and export the query database
0.1.1 13 Feb 2026 1.38.1 / 0.5.0 Cost-based planner by default, streaming shuffles for local nodes, streaming as-of joins, and distributed stable sort, top-k and distinct

Access the released artifacts

Helm chart

Helm chart releases are announced and documented on our chart repo. The summary of each chart release lists its cluster software version. These commands list it too:

helm repo add polars-inc https://polars-inc.github.io/helm-charts
helm search repo polars-inc --versions

The container images are hosted on Docker Hub and are tagged with the versions listed above.

Bare-metal package

The polars-nanook-compute package is published on PyPI, and Get started on bare-metal installs it. An air-gapped cluster installs the same package from PyPI or your organization's package mirror, as described in Install and license. Package versions match the cluster software releases above.