Daemon (buckd)

The first time that a Buck2 command is run, Buck2 starts a daemon process for the current project. For subsequent commands, Buck2 checks for the running daemon process and, if found, uses the daemon to execute the command. Using the Buck2 daemon can save significant time as it enables Buck to share cache between Buck2 invocations.

By default, there is 1 daemon per project root, you can run multiple daemons in the same project by specifying an isolation dir.

While it runs, the Buck daemon process monitors the project's file system for changes. The Buck daemon excludes from monitoring any subtrees of the project file system that are specified in the [project].ignore setting of .buckconfig.

You can see detailed information about the status of the daemon by running buck2 status.

Experimental daemon sharing

Daemon sharing is opt-in and off by default. BUCK2_SHARE_DAEMON=true lets tenants in different isolation directories share a daemon while retaining separate output and cache state.

In this mode buck2 kill evicts the selected (project root, isolation directory) tenant from the daemon and cancels its work. Eviction waits for cancellation and cleanup before returning. A later command can recreate the tenant only after the old generation releases its output and cache directories. Other tenants continue running; the daemon shuts down after its final tenant is evicted. buck2 clean uses the same ownership lock while deleting the selected tenant's state, and buck2 killall --repo evicts the repository's tenants instead of killing processes. Incompatible startup settings or binaries are reported without restarting the daemon; stop the affected daemon's tenants explicitly before changing these settings.

To also share across checkouts, set BUCK2_DAEMON_GROUP to the same name in each checkout. A group name must be a single valid filename component. Each tenant still uses its own checkout's files, configuration and outputs, so checkouts may stay on different commits.

Daemon groups require compatible startup configuration and binary versions. An incompatible invocation fails without restarting the group's daemon. Use different group names for A/B instances with different startup settings, or evict all tenants in a group before changing those settings.

[buck2] share_dice = false (the default) creates a separate DICE engine for each tenant. Setting it to true shares one engine within the daemon. Each tenant keeps its own root, repository data, and command coordinator. Matching tracked inputs allow results to be reused across tenant roots. File reads depend on the watcher's filesystem baseline: EdenFS and scm-aware Watchman can use a shared mergebase when they have recorded the checkout's differences from it; watchers without that guarantee use a unique baseline. Artifact filesystem state also identifies the checkout's project root.

The client selects this setting at daemon startup. It is part of the daemon configuration checked on connection. Changing it restarts an unshared daemon. A shared daemon reports the mismatch instead of restarting, so stop its tenants with buck2 killall --repo first; for a named group, choose a new group or evict its existing tenants. Command event logs include a dice-sharing:true or dice-sharing:false tag.

Cycle detection is always enabled in both modes. Shared engines require compatible paging, digest configuration, and Buck settings. Paging uses the initial tenant's DICE storage path for the lifetime of the daemon. Evicting a tenant releases its roots after admitted commands finish; other tenants keep running on the shared engine.

Killing or disabling the Buck daemon

The Buck daemon process is killed if buck2 clean or buck2 kill commands are run. Note that they won't kill the daemon associated with custom isolation dirs. To do that, run using the --isolation-dir option (buck2 --isolation-dir <dir> <command>)