Batch Jobs
Queued work for evaluations, document processing, data preparation, and finite agent tasks.
Create and monitor jobs in the console or Control API. New jobs prefer Europe (eu-1); US East (us-1) remains available when explicitly selected. Execution requires an accepted compute quote.
Run finite workloads in Europe or US East.
The contract defines CreateJob, job reads, run listing, run reads, cancellation, resources, placement and accepted compute quotes. Jobs run in isolated runtimes in US East. Use a digest-pinned Linux/amd64 image compatible with a non-root, read-only filesystem. Omitted placement selects us-1; creation requires an available, accepted quote.
Start in the console
Open Batch Jobs, choose an image and command, review the exact price, and create the job. The console supports a 1-vCPU, 1-GiB size; the API accepts other supported sizes. Cold starts can take several minutes. Review run status, cancel running work, and download retained output.
Use your organization's private image
Build locally and push to Tarka's registry, then supply registry.tarka.rest/orgs/<organization UUID>/<repository>@sha256:<digest>. Docker prints the manifest digest after a successful push. The job must belong to that organization; Tarka supplies pull access automatically. Public images from supported registries remain supported.
Job images are limited to 512 MiB compressed and 2 GiB expanded. They run as UID 65532 with a read-only root filesystem; use the job's writable scratch space for temporary files. Tags alone are not accepted.
Consent before execution
Review the compute quote and usage policy before creating work. The US East contract uses one on-demand attempt by default. Queue and setup time are free; started runtime has a minimum charge. Customer application failures and cancellation pay for started runtime; confirmed platform failures waive the operation, including network usage.
Cancellation is a request to stop. Keep the run ID and read its final state before treating work as finished. Ambiguous execution must not be replayed automatically.