Fix the registered config name for leaderbalancernodemutetimeout. by @WillemKauf in #31360
Fixes a bug in which L0 batches in a cloud topic forgot to preserve lastoffsetdelta in their header, leading to an under-declared last offset which can stall consumers, skip records, or halt exact-offset replication. by @WillemKauf in #3136
Fixes a bug in which topics with min.compaction.lag.ms left unconfigured with produced batches holding timestamps in the future would be considered ineligible for compaction by @WillemKauf in #31459
Fixes a bug in which transient TOPICAUTHORIZATIONFAILED errors and SASL authentication failures were possible during application of a controller snapshot. by @WillemKauf in #31437
Hi, I'm not sure where the Dockerfile lives and this might be the wrong repo but I'll throw the issue here in the hope that it's the best place. :) ## The goal (aka what we're trying to do): We are trying to set up a [docker-compose example with redpanda and tremor](https://github.com/tremor-rs/tremor-redpanda/) to make it easy for users to test out this int
### Who is this for and what problem do they have today? - Developer wants to interact with redpanda. The developer is not involved in Redpanda's hosting or installation. - Developer uses Linux distributions other than rhel/centos/fedora/ubuntu. ### What are the success criteria? - rpk can be installed with a shell one-liner on "all" Linux Distros and OSX -
## What happened `rpk connect upgrade` fails for every currently installed Redpanda Connect version, because it parses the *current* version through a regex that caps every segment at two digits before comparing it to latest. Connect's minor version passed 100 in 2026, so upgrade is broken for every install since ~mid-2026. ``` $ rpk connect --version Versio
The sysroot's `linux-libc-dev` decides which kernel uapi constants the build can see, no matter which kernel the broker actually runs on. Ours is built from Ubuntu 22.04, which pins that at 5.15 and costs us every addition since October 2021. Two of those we already reach for and do not find, and in both cases someone worked around it locally rather than the
Backport https://github.com/redpanda-data/redpanda/issues/31545 to branch v26.2.x. Requested by PR https://github.com/redpanda-data/redpanda/pull/31546
Backport https://github.com/redpanda-data/redpanda/issues/31545 to branch v26.1.x. Requested by PR https://github.com/redpanda-data/redpanda/pull/31546
Backport https://github.com/redpanda-data/redpanda/issues/31545 to branch v25.3.x. Requested by PR https://github.com/redpanda-data/redpanda/pull/31546
Seastar's CMake build defines `SEASTAR_SHUFFLE_TASK_QUEUE` in the Debug and Sanitize configurations, randomizing task queue execution order to surface latent task ordering assumptions. The Bazel debug build was missing this define; add it to restore the old CMake behavior. ## Backports Required - [x] none - not a bug fix - [ ] none - this is a backport - [ ]
Schema Registry API sync only propagates changes on its periodic full sync, so a registration can wait a whole interval and every propagation costs a whole-registry scan of both sides. This adds a change feed: the link consumes the source's _schemas topic over the Kafka API and re-reads only the targets whose records it saw, reconciling them by the same per-
## Problem On the RPC audit sink (the default since v25.3), `create_internal_topic()` is the **entire** configure step of `audit_client::initialize()`, and it unconditionally sends `CreateTopics` to the controller — so every audit enable/startup requires a reachable controller leader, even when the audit topic already exists. `topics_frontend::autocreate_top
Fixes #31545 ## What this does `VersionFromString` (`pkg/redpanda/version.go`) parsed a version string through a regex that capped every segment at two digits: ```go regexp.MustCompile(`^v?(\d{1,2})\.(\d{1,2})\.(\d{1,2})(?:\s|-rc\d{1,2}|-dev|-nightly|$)`) ``` It's called by `connect/upgrade.go`'s `connectVersion` helper to parse the *currently installed* Con
`NodeDecommissionWaiter.wait_for_removal` asserts that a decommission has "stopped making progress" if `replicas_left`/bytes-left-to-move hasn't improved within `progress_timeout`. In tests that run `FailureInjectorBackgroundThread` alongside node operations (e.g. `ShadowLinkingRandomOpsTest.test_node_operations`), this assertion can trip even though the sys
Backs `crc::crc32c` with abseil's CRC32C instead of google/crc32c. The wrapper becomes `basic_crc32c<Backend>` with backends for both libraries, and `crc::crc32c` is an alias for the abseil instantiation, so production switches over with no call site changes. CRC32C values are identical between the backends, which is what makes the switch safe: they are pers
## Summary Copyedits surfaced by auditing the docs team's rpk description overrides (`docs-data/rpk-overrides.json` in redpanda-data/docs) against rpk's source text — these are cases where the docs override exists only to paper over a source defect, so fixing the source lets the docs drop the override: - `profile clear`: "unset **an prod** cluster profile" →
## What Removes the one-shot IMDS connect probe from the cloud instance-type detector (`instance_type_detector.cc`). ## Why The detector probed the EC2 IMDS endpoint with a single bounded `base_transport` connect before issuing any request. That was a workaround (added alongside the capacity-metrics feature in #30742) for `http::client::get_connected` retryi
Revives #8570 (rebased onto dev, see #8565 for the original motivation): put the seastar memory diagnostics dump, which is emitted at ERROR level when a node runs out of memory, on the default log allow list so an OOM surfaces as the "redpanda crashed" diagnostic rather than a BadLogLines about the dump. Also clarifies in the raise_on_bad_logs docstrings tha
Adds `--remote_download_minimal` in bazelrc that stops CI materialising outputs no CI step reads. This reduces downloaded and materialized-on-disk bytes by about 70% and shaves maybe ~1 minute off of a cached build. jira: [CORE-16997](https://redpandadata.atlassian.net/browse/CORE-16997) ## Activated later It is inert until something passes `--config=ci-remo
Schema Registry brokers abort when a request fails before its deferred authorization check. With `schema_registry_enable_authorization` set, a few endpoints only learn which resource to check against partway through handling, so the check runs inside the handler. Three individually reasonable things interact badly: 1. `request_auth_result` protects against h
`segment_concatenation_test` is declared `timeout = "short"` (60s), but the two 100-segment params take ~11s each locally, putting the whole target at 36s — 61% of its budget. The test is fsync-bound (`fsync::yes` appends plus sanitized file config), so on a CI host with slower syncs it runs past 60s and bazel kills it mid-test, which reads as a flake. Bumpi
Adds a `--cluster-type` flag to `rpk generate grafana-dashboard` (K8S-855, https://redpandadata.atlassian.net/browse/K8S-855). The default value (`default`) leaves the existing `--dashboard` flow untouched; `stretch` swaps in dashboard variants for stretch clusters. The stretch variant of the `operations` dashboard is a new dashboard focused on stretch clust
Units released back to their originating shard are returned via an unawaited cross-shard call, so they are not guaranteed to be visible to the immediately following read; under `--@seastar//:shuffle_task_queue=true` the read can run first (reproduced in 12/30 runs). Poll for the release instead. Test-only change. ## Backports Required - [ ] none - papercut/n
<!-- See https://github.com/redpanda-data/redpanda/blob/dev/CONTRIBUTING.md#pull-request-body for more details and examples of what is expected in a PR body. Content in this top section is REQUIRED. Describe, in plain language, the motivation behind the change (bug fix, feature, improvement) in this PR and how the included commits address it. Add the GitHub
The Redpanda Load Factor dashboard was recently added to the [observability repo](https://github.com/redpanda-data/observability) in redpanda-data/observability#51. It shows utilization relative to capacity (load factor) for the key broker resources: CPU (reactor), IO scheduler, disk IOPS, memory, network bandwidth, and client connections, so users can ident
CORE-16998, CORE-17000, CORE-17001 Three bugs in the classic group's offset accounting, all the same defect: live state and a replay of the log resolve the same key differently. The group's committed offset then changes on a leader change, a restart or compaction, and no client asked for it. None is a KIP-848 regression, and each commit carries its own unit
The multi-term segments design assumes raft_configuration batches demarcate term boundaries. They do not. A candidate that wins an election but fails the local append of its initial configuration batch stays leader (`vote_stm.cc`: "even if failed to replicate, don't step down"). Nothing retries the append. Data of the new term then commits with no configurat
## Summary - `test_producer_ids_failover` occasionally hit kgo-verifier's "possible idempotency bug" check right after `failover_link()` — failover completion is a health-report heuristic (`link_status_reconciler::try_finish_failover`), not a guarantee that every previously-mirrored write has landed, so a fresh producer's one-shot `list-offsets(-1)` snapshot
Two raft liveness gaps let a follower storm elections against a live leader, the family of symptoms previously reported in #30815: * A follower in active recovery rejects lightweight heartbeats without refreshing its election timer, and the full heartbeat fallback after a lightweight heartbeat failure can be suppressed indefinitely by stalled in-flight appen