warren.storage

A requested resource does not exist in the store.

A resource with the given identity already exists.

A store operation failed for a transient, retryable reason.

Signals a temporary backend condition (connection blip, failover, server-selection timeout) rather than a permanent fault. Callers that cannot tolerate a dropped operation (e.g. the JobStatusWorker observer, where a lost write strands a job) can retry on it without depending on the concrete backend.

The job store and the job-results store raise this type for transient failures, and other errors from these two stores are permanent. The document, results and publishing stores raise the backend error unchanged.

Run operation, retrying TransientStoreError with exponential backoff.

Permanent errors propagate immediately — only TransientStoreError is retried. After attempts transient failures the last error is re-raised, leaving the caller to decide whether to propagate or drop.

Defaults (5 attempts, 1s base, 8s cap → ~15s total) are sized to ride a MongoDB replica-set failover/election (electionTimeoutMillis defaults to 10s); a longer outage is treated as a real one.

Raises:

• TransientStoreError: if every attempt fails transiently.

Re-raise transient pymongo errors from func as TransientStoreError.

Permanent errors propagate unchanged. Wrap async store methods so callers can retry on the standard TransientStoreError type.

Class decorator: apply classify_transient to every public async method.

It wraps only the methods that the class itself defines, not inherited methods.

Applies the transient-error contract store-wide, so the whole implementation honours "transient backend failures are raised as TransientStoreError" without per-method boilerplate. Private (_-prefixed) methods are left untouched — public methods that call them still classify at the public boundary.

The scope of the current handler, set by the consumer for each message.

Return the raw scope of the current handler.

Whether a raw scope is a valid namespace.

A valid scope is a string of 1 to 40 characters that matches [a-z0-9][a-z0-9-]{0,39}.

Fields for a control row, without rejecting invalid envelopes.

Resolve content collections and prepare indexes once per namespace.

Calling the instance returns the database prefix + scope for the scope of the current handler.

Raises:

• HardFailureException: On call, if the scope is malformed, or missing with no fallback database.

Erase scoped content, cache keys and control rows after traffic is stopped.

It deletes three things:

• the scope database prefix + scope (or each of its collections, if the user cannot drop the database)
• the Redis keys that match s:{scope}:*
• the rows with {"scope": scope} in each collection of control_database

Raises:

• HardFailureException: If the scope is malformed.
• ValueError: If the scope database equals control_database, or a MongoDB or Redis client is missing.

Document store with read-through/write-through cache.

Wraps any DocumentStoreInterface with a CacheInterface[Dict] for fast reads. All writes go to both store and cache (write-through). Reads check cache first, falling back to store on miss (read-through).

Cache failures are logged as warnings and do not propagate, except HardFailureException (a missing or malformed scope on a scoped cache). The underlying store is the source of truth.

Hands out in-process stores so co-hosted runners share state.

A database shares state by name: two workers that open the collection "chunks" see the same documents. In one process that role is played by this registry. Document stores, results stores and caches are get-or-create on a name, so the worker that writes "chunks" and the worker that reads it get the same store. The job store and the job-results store are one per registry. Build one per process (or one per test, for a fresh world) and pass it to whatever constructs the runners.

The first call for a name fixes that store's configuration; later calls return the existing store and ignore their keyword arguments.

Configure a positive TTL; leave indexes untouched when unset.

Drops an existing single-field index on field with other options, then creates the TTL index.

Raises:

• ValueError: If seconds is not positive.