4.4 Waiting and Contention
4.4.1 Request deadlines #
Each RSI operation uses the absolute CLOCK_MONOTONIC deadline supplied by
LCS through REG_SRC_REQUEST_CONTEXT. The budget includes kernel admission,
queue time, connection-pool waits, publication-gate waits and SQL statements;
receiving the request does not restart it. A contended wait that expires
returns RSI_TXN_BUSY; successful operations retain their success status.
SQLite's busy timeout is zero because its native busy wait can outlast a
canceled Go context. Busy classification uses SQLite and Go error codes.
The lifetime extension is negotiated before hive registration. loregd requires version 1 and fails startup on a kernel without it. The 25-second local budget is only a fallback for direct, unnegotiated test clients.
4.4.2 A transaction holding the writer #
A competing mutation returns RSI_TXN_BUSY immediately while a read-write
transaction owns the hive's write connection. This includes non-transactional
key drops, conditional value writes, flushes and layer deletion. Admission
and binding are coordinated under the hive gate, so the check cannot race a
new binding. A read-only snapshot is not counted as a writer.
Both stores use WAL with private caches. Ordinary reads see committed data while a writer remains open; backup snapshots do not hold volatile table locks that block writers. A short per-hive gate excludes reads from commit publication and non-transactional mutations, not from an entire transaction's lifetime. Different hives have separate gates.
4.4.3 Cancellation and cleanup #
Operations on the same transaction are serialized, including commit and abort. Rollback uses a fresh cleanup context, and an uncertain connection is discarded rather than returned to the pool with a live transaction. Every replacement physical connection receives its WAL, foreign-key and volatile-attachment setup again.
SQLite may automatically roll back an interrupted write. Such a bound transaction is marked failed and subsequent use returns a storage error; LCS must abort it. It is never silently rebound or allowed to autocommit.
LCS owns transaction lifetime. A live transaction file object (including all duplicated handles) or an in-kernel operation owns the shared kernel transaction record. A source can renew its five-second cleanup lease only while that owner and source session remain valid. SQL activity neither creates ownership nor extends a transaction's lifetime deadline. An internal operation may have no separate lifetime limit; its live operation scope is still required.
A separate control loop renews leases with three seconds or less remaining, checks expiration every 100 milliseconds, and reclaims expired transactions. Control ioctls consume no ordinary RSI request slots. Final owner release, abort, timeout or a lost source session prevents renewal even if an abort message cannot be delivered. Expired transactions reject new operations; late renewals cannot revive them. Cleanup serializes on the transaction's operation gate and waits for an already-admitted commit to resolve. A request timeout is not evidence that a commit failed. The request deadline gates commit admission; once admitted, SQL COMMIT runs without caller cancellation and its actual result is reported, including through normal late-response recovery. Shutdown drains requests and cleanup before closing databases.