# etcd v3.7.4

2026-11-19 · [Original release notes](https://github.com/etcd-io/etcd/releases/tag/v3.7.4) · [Full summary](https://ratatosk.io/en/releases/etcd/v3.7.4)

Breaking changes 1 · Important fixes 2 · Other changes 6

## Check before upgrading

### Upgrade guide

- **Upgrade guidance for etcd v3.7.4**
  > Please check out [CHANGELOG](https://github.com/etcd-io/etcd/blob/main/CHANGELOG/CHANGELOG-3.7.md) for a full list of changes. And make sure to read [upgrade guide](https://etcd.io/docs/v3.7/upgrades/upgrade_3_7/) before upgrading etcd (there may be breaking changes).

  - Zero-downtime etcd v3.6-to-v3.7 upgrade guidance
  - Guide to preparing for an upgrade
  - Updated 3.6 members required for a 3.7 upgrade: Before upgrading, all cluster members must be on 3.6.11 or later; earlier 3.6 patch releases may not be compatible with a rolling upgrade to 3.7.
  - v2 data migration guide for older releases
  - CHANGELOG covering upgrades for etcd embedders, `clientv3` users, and internal-package dependents
  - v3.5-to-v3.6 guide to removed flag replacements
  - v3.7 upgrade requires etcd v3.6.11 or later: Clusters on versions older than v3.6 must first be upgraded to v3.6; etcd supports upgrading only one minor version at a time.
  - Healthy-cluster prerequisite for rolling upgrades: A healthy cluster is required for a smooth rolling upgrade; the note names `etcdctl endpoint health` as the health check.
  - Staging validation for etcd-dependent services: The note places these tests in staging, before the upgrade is deployed to production.
  - Snapshot backup for etcd upgrade rollback: A snapshot backup taken before the upgrade can be used to roll back to the existing etcd version if the upgrade goes wrong.
  - Upgrade guide covers cluster snapshots and rollback
  - Binary rollback unavailable after all members reach v3.7: Recovery options are restoring the pre-upgrade snapshot or following the official downgrade guide.
  - Cluster health and version checks: The procedure uses `etcdctl ... endpoint health` and `curl .../version` to check that the cluster is healthy and running v3.6.11 or later. The sample output shows all three endpoints healthy, with etcdserver 3.6.12, etcdcluster 3.6.0, and storage 3.6.0.
  - Leader snapshot backup for downgrades
  - Upgrade procedure for stopping an existing etcd server: Other members may report errors while the server is temporarily disconnected. If it is the leader, it transfers leadership before exiting.
  - Upgrade guide covers endpoint status checks for members and the cluster

## Breaking changes (1)

### Deprecated etcd `--experimental-*` startup flags removed

- Affects: etcd deployments whose startup configuration sets any of the listed `--experimental-*` flags
- Action required: Replace any listed `--experimental-*` flags with their v3.6 equivalents before rolling to v3.7.

In v3.6, each deprecated flag was replaced by a non-experimental flag of the same name or a `--feature-gates` entry. In v3.7, etcd fails to start if any removed flag remains configured.

> All deprecated --experimental-* flags have been removed in v3.7 ( #19959 ). In v3.6 each of these was replaced either by a non-experimental flag of the same name or by a --feature-gates entry. If you still have any of these set, replace them with the v3.6 equivalent before rolling to v3.7, otherwise the v3.7 process will fail to start.
> -etcd --experimental-bootstrap-defrag-threshold-megabytes
> -etcd --experimental-compact-hash-check-enabled
> -etcd --experimental-compact-hash-check-time
> -etcd --experimental-compaction-batch-limit
> -etcd --experimental-compaction-sleep-interval
> -etcd --experimental-corrupt-check-time
> -etcd --experimental-distributed-tracing-address
> -etcd --experimental-distributed-tracing-instance-id
> -etcd --experimental-distributed-tracing-sampling-rate
> -etcd --experimental-distributed-tracing-service-name
> -etcd --experimental-downgrade-check-time
> -etcd --experimental-enable-distributed-tracing
> -etcd --experimental-enable-lease-checkpoint
> -etcd --experimental-enable-lease-checkpoint-persist
> -etcd --experimental-initial-corrupt-check
> -etcd --experimental-memory-mlock
> -etcd --experimental-peer-skip-client-san-verification
> -etcd --experimental-snapshot-catchup-entries
> -etcd --experimental-stop-grpc-service-on-defrag
> -etcd --experimental-txn-mode-write-with-shared-buffer
> -etcd --experimental-warning-apply-duration
> -etcd --experimental-warning-unary-request-duration
> -etcd --experimental-watch-progress-notify-interval
> Flags removed

https://etcd.io/docs/v3.7/upgrades/upgrade_3_7/

## Important fixes (2)

### `--force-new-cluster` bootstrap rejects commit indexes beyond the WAL

- Affects: members using `--force-new-cluster`

When the backend consistent index was ahead of the last WAL entry, bootstrap could panic and permanently brick the member. It now returns an error instead of proceeding with an adjusted commit index beyond the last WAL entry.

> - Fix [`--force-new-cluster` permanently bricking a member whose backend consistent index is ahead of its WAL](https://github.com/etcd-io/etcd/pull/22685)

https://github.com/etcd-io/etcd/blob/main/CHANGELOG/CHANGELOG-3.7.md

### Legacy v2 WAL replay panic fixed

- Affects: etcd members with legacy v2 member-attribute or cluster-version metadata entries in their WALs

Startup replay could panic on legacy v2 member-attribute entries. Replay now drops legacy v2 member-attribute and cluster-version metadata entries, republishes member attributes before serving requests, and newer members maintain the cluster version.

> - Fix [etcd panics on startup replaying legacy v2 member-attribute WAL entries](https://github.com/etcd-io/etcd/pull/22804)

https://github.com/etcd-io/etcd/blob/main/CHANGELOG/CHANGELOG-3.7.md

## Other changes (6)

- Race condition in `alarmStore` access fixed
- etcd v2 store and related capabilities removed
- etcd's protobuf code migrates to `google.golang.org/protobuf`
- gRPC proxy logging and tags switch to `go-grpc-middleware` v2
- etcd gRPC tracing migrates to `NewServerHandler`
- Mixed-version etcd cluster upgrade support

---

Summaries are produced by AI, so they can be wrong, and they are provided without warranty.
