mirror of
https://github.com/BerriAI/litellm.git
synced 2026-10-05 02:41:56 +00:00
8b63b7045c
Background interaction billing previously lived only in the creating pod's memory: a DELETE routed to another pod or a proxy restart meant the completed provider work was never billed and the budget reservation was refunded at the poll timeout. The poll task now persists a settlement context row (virtual key attribution, model, original request id, budget reservation) in a new LiteLLM_BackgroundInteractionSettlementTable so any pod can settle. A conditional update flipping status pending to settled is the cross-pod exactly-once claim; every outcome (bill, release, timeout give-up) must win it before touching spend counters, and a losing pod finalizes its local reservation copy without touching shared counters. The delete route settles registry misses through the store by rebuilding a billable logging context, resolving provider credentials through the router by model group. A per-pod sweep job resumes pending settlements after restarts and surfaces timed-out entries as abandoned rows instead of silently reconciling to zero. Without a registered store (SDK usage) behavior is unchanged. |
||
|---|---|---|
| .. | ||
| migrations | ||
| __init__.py | ||
| _logging.py | ||
| schema.prisma | ||
| utils.py | ||