Documentation menu
Documentation

Bazel Missing digest from the remote cache

A cached action result can refer to output bytes that the cache can no longer serve. Identify which blob request failed, check the backend and Bazel version, then verify that a fresh consumer can download the outputs.

Why a result can exist while its output is missing

The error may contain one or both of these fragments.

Failed to fetch blobs because they do not exist remotely
Missing digest

Bazel stores action-result metadata in the action cache and output bytes in the content-addressable store, or CAS. The result identifies content by digest. An unavailable blob can prevent Bazel from materializing an output even when it has already found result metadata. A client may also remember metadata from an earlier build after the backend has evicted the referenced content.

Check the failing operation. A missing HTTP /ac/ entry is an ordinary cache miss. A missing /cas/ blob referenced by a result means the content needed for that result is unavailable. gRPC and remote-execution failures can involve missing inputs too. The Bazel caching reference explains the two stores.

Establish the version and backend

bazel info release

Record whether the build uses an HTTP cache, a gRPC cache, or remote execution, and whether the failure followed a long idle period or backend eviction. Ask the cache operator to check the failed operation, missing-blob response, and retention behavior using metadata. These details guide recovery without downloading or sharing private artifact contents.

Upstream reports document CacheNotFoundException and Missing digest with Bazel 6.3.1 and 6.4.0rc1 using gRPC, remote execution, and deferred output downloads. That report was closed with an upstream fix. Check the release notes for your pinned version before applying a version-specific retry workaround. See the upstream eviction report and fix discussion.

Rebuild the affected result from trusted inputs

Restore backend health first. Have a trusted producer recompute the affected target and publish its outputs using the intended write credentials. If the same digest repeatedly disappears, investigate retention and the availability of referenced blobs with the operator. Repeated retries alone cannot supply content that remains unavailable.

For a project that supports local execution, this isolated comparison uses a fresh output base with disk caching, remote caching, and remote execution disabled. Replace the target and keep the source revision and toolchain the same. It creates new local state, so expect a full build of the selected target.

cache_check_base="$(mktemp -d)"
bazel --output_base="$cache_check_base" build //path:target \
  --disk_cache= \
  --remote_cache= \
  --remote_executor=

This comparison helps establish whether the target can build using the local toolchain. A project configured with remote-only execution strategies needs a compatible local strategy and toolchain before using it.

After the trusted producer has published a complete result, use another fresh consumer output base with your normal remote-cache configuration. Keep the disk cache disabled for the check and request the outputs your next step needs. Confirm a remote cache hit, a successful download, and the expected files. The Bazel remote cache guide and Cachely HTTP setup cover that workflow.

Scope of the explanation

Sources checked November 11, 2026. The explanation follows Bazel's protocol documentation and the versioned upstream report above. The diagnostic options were checked against Bazel 8.1.0 option definitions. Recovery and automatic retry behavior depend on the Bazel version, backend protocol, and execution mode.