Choose the operating arrangement
| Starting point | What you take on | Where Hermetiq fits |
|---|---|---|
| Managed cache and execution | Agree supported workloads, access, hosting, support, and the responsibilities you retain. | Managed Cache & Execution supplies Buildbarn-based infrastructure with analytics and AI included. |
| A new cluster you operate | Provision storage and services; configure security, workers, monitoring, upgrades, and incident response. | Use the public Buildbarn Helm chart for Kubernetes and connect analytics when needed. |
| Buildbarn already in production | Keep operating your existing endpoints, worker pools, storage, and configuration. | Connect Build Analytics & AI and the telemetry needed for your investigation. |
Hermetiq Cloud and a deployment in your own environment are hosting choices. They do not settle who owns upgrades, capacity, or incidents. Put those responsibilities in the deployment scope. The open-source chart is a deployment tool; using it alone does not purchase managed operations.
To evaluate without operating a cluster, create a Hermetiq Cloud project and use Quickstart’s shared-worker execution settings. No separate executor-provisioning request is needed for that trial. Shared Ubuntu 22.04 workers have limited idle capacity; custom platforms and customer-hosted deployments need separate planning.
Separate the cache from the workers
The content-addressable store (CAS) holds blobs such as inputs and outputs. The action cache records results for action digests. Remote execution additionally needs scheduling, workers, and an execution environment with the required tools.
For an action that executes remotely, Bazel makes inputs available, execution work is scheduled for a compatible worker, and the worker fetches inputs, runs the command, and uploads outputs. A reachable cache endpoint does not prove that a worker can run your action.
- Bazel client
- Storage and scheduler services
- Compatible worker
- Outputs and action result
Prepare a workload and an execution platform
Start with one supported Bazel target and a test. Record the client configuration, intended instance name, execution-platform properties, worker image, and toolchain versions. The worker pool must satisfy the platform requested by the action; spare capacity in a different pool may be irrelevant.
For Kubernetes, follow the chart's current requirements and values instead of copying image tags or install commands from an old article. Review storage persistence, networking, TLS, authentication, authorization, resource limits, and the worker execution boundary for your environment. The upstream bb-deployments examples are useful for learning the topology; they are not a complete production security or operations plan.
Verify execution, then verify reuse
- Build a representative target with the intended remote configuration. Inspect Bazel execution evidence and corresponding server records to confirm an action actually executed remotely.
- Use a separate checkout or isolated output base with equivalent inputs and configuration to test remote-cache reuse. An unchanged local rerun can be satisfied locally without contacting the remote cache.
- Run the workload's tests and compare outputs. Inspect failures, not only the command's exit code.
- Preserve the version and configuration revision, invocation ID, platform, and relevant action identifiers so a later regression has a baseline.
Use the endpoint, instance, authentication helper, and platform supplied for your deployment. There is no universal remote-executor command that is correct for every Buildbarn cluster.
Check an operation change against build evidence
The Cloud pilot verifies real execution and cache reuse on Hermetiq’s shared Buildbarn workers. For an existing cluster, the recorded queue investigation follows a different question: whether the worker’s configured concurrency is limiting a waiting workload.
We rechecked that September 8, 2026 experiment against all six saved runs, 102 action identities and results, the exact concurrency change from 1 to 4, and 66 file checksums. The median of the per-run mean queue delays fell from 7.893954 to 1.809266 seconds. One post-change run still had worse end-to-end time than every baseline run. Queue relief does not by itself establish a faster build. These are controlled local results, not a Cloud performance claim.
Plan the work after installation
| Responsibility | A useful operating question |
|---|---|
| Queues and capacity | Which instance and platform are waiting, and do compatible workers have spare capacity? |
| Storage | Are action results and referenced blobs available for the reuse window the workload needs? |
| Changes | Can you identify the image, configuration, or toolchain revision that introduced a regression and roll it back? |
| Reliability | Who responds when cache requests fail, workers disconnect, or input and output transfers stall? |
| Access and cost | Who controls client credentials, tenant boundaries, retention, worker sizing, and infrastructure spend? |
Hermetiq's monitoring guide follows these signals through a build. The recorded queue experiment shows one capacity investigation; it is an example to inspect, not a sizing formula for another cluster.
Sources and review
Reviewed against the primary references below and Hermetiq's setup and deployment documentation on November 19, 2026. Check flags and requirements against the versions you run. The adoption and operating checklists are engineering guidance, not measured performance claims.