Revert "fix(proxy): recover from prisma-query-engine zombie process"

[yuneng-jiang]

Reverts #21707

Issue:

Loops infinitely, and does not allow me to access the DB locally

14:19:40 - LiteLLM Proxy:INFO: utils.py:3859 - Found prisma-query-engine at PID 45601.
2026-02-21 14:19:40 - LiteLLM Proxy - INFO - utils.py:3859 - _start_engine_watcher() - Found prisma-query-engine at PID 45601.
14:19:40 - LiteLLM Proxy:INFO: utils.py:3877 - Watching engine PID 45601 via /proc polling (1s fallback).
2026-02-21 14:19:40 - LiteLLM Proxy - INFO - utils.py:3877 - _start_engine_watcher() - Watching engine PID 45601 via /proc polling (1s fallback).
14:19:40 - LiteLLM Proxy:INFO: utils.py:3995 - Prisma DB reconnect succeeded. reason=engine_process_death
2026-02-21 14:19:40 - LiteLLM Proxy - INFO - utils.py:3995 - _attempt_reconnect_inside_lock() - Prisma DB reconnect succeeded. reason=engine_process_death
14:19:40 - LiteLLM Proxy:ERROR: utils.py:3802 - prisma-query-engine PID 45601 disappeared; triggering reconnect.
2026-02-21 14:19:40 - LiteLLM Proxy - ERROR - utils.py:3802 - _poll_engine_proc() - prisma-query-engine PID 45601 disappeared; triggering reconnect.
14:19:40 - LiteLLM Proxy:WARNING: utils.py:3987 - Attempting Prisma DB reconnect. reason=engine_process_death
2026-02-21 14:19:40 - LiteLLM Proxy - WARNING - utils.py:3987 - _attempt_reconnect_inside_lock() - Attempting Prisma DB reconnect. reason=engine_process_death
14:19:40 - LiteLLM Proxy:WARNING: utils.py:3915 - prisma-query-engine PID 0 is dead; performing heavy reconnect.
2026-02-21 14:19:40 - LiteLLM Proxy - WARNING - utils.py:3915 - _run_reconnect_cycle() - prisma-query-engine PID 0 is dead; performing heavy reconnect.
query-engine ac9d7041ed77bcc8a8dbd2ab6616b39013829574

@hcavarsan Please take a look when you get a chance

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Error		Feb 21, 2026 10:21pm

[greptile-apps]

Greptile Summary

This PR reverts #21707, which added a prisma-query-engine process watchdog (SIGCHLD handler, pidfd monitoring, /proc polling, and heavy reconnect logic). The reverted feature was causing immediate false-positive engine death detection on startup — the watcher found the engine PID, then immediately concluded it had disappeared and triggered an unnecessary reconnect cycle (visible in the logs as "PID disappeared" right after "Found prisma-query-engine at PID").

The revert:

• Removes ~350 lines of engine PID tracking, SIGCHLD signal handling, pidfd_open monitoring, /proc polling, zombie reaping, and heavy reconnect (via recreate_prisma_client) from PrismaClient in litellm/proxy/utils.py
• Deletes the corresponding 431-line test file tests/litellm/proxy/test_prisma_engine_watchdog.py
• Restores the simpler _run_reconnect_cycle that only does lightweight disconnect → connect → SELECT 1
• Preserves the existing DB health watchdog (_db_health_watchdog_loop) and all existing self-heal tests in test_prisma_self_heal.py

The revert is clean: no dangling references to removed methods or import signal remain in the codebase.

Confidence Score: 5/5

• This PR is safe to merge — it is a clean revert that removes a buggy feature and restores well-tested prior behavior.
• This is a straightforward git revert of a single feature PR (fix(proxy): recover from prisma-query-engine zombie process #21707). The revert cleanly removes all added code and tests with no residual references. The reverted-to state has existing test coverage via test_prisma_self_heal.py. The PR includes clear evidence of the bug being fixed (log screenshots showing false-positive engine death detection).
• No files require special attention.

Important Files Changed

Filename	Overview
litellm/proxy/utils.py	Reverts the engine watchdog feature: removes ~350 lines of engine PID tracking, SIGCHLD handling, pidfd monitoring, /proc polling, and heavy reconnect logic. Restores the simpler lightweight reconnect cycle (disconnect → connect → SELECT 1). No residual references to removed methods or imports.
tests/litellm/proxy/test_prisma_engine_watchdog.py	Deletes the entire 431-line test file for the reverted engine watchdog feature. The remaining self-heal tests in tests/test_litellm/proxy/db/test_prisma_self_heal.py still cover the lightweight _run_reconnect_cycle and watchdog start/stop lifecycle.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["DB Health Watchdog Loop<br/>(periodic SELECT 1 probe)"] -->|"Probe fails"| B["attempt_db_reconnect()"]
    B --> C["_run_reconnect_cycle()"]
    
    subgraph BEFORE["Before Revert (PR #21707)"]
        D["_start_engine_watcher()"] --> E{"Engine PID alive?"}
        E -->|"Dead (SIGCHLD/pidfd/poll)"| F["Heavy reconnect:<br/>recreate_prisma_client()"]
        F --> D
        E -->|"Alive"| G["Lightweight reconnect:<br/>disconnect → connect → SELECT 1"]
    end
    
    subgraph AFTER["After Revert (This PR)"]
        H["_run_reconnect_cycle()"] --> I["Lightweight reconnect:<br/>disconnect → connect → SELECT 1"]
    end
    
    C -.->|"Removed"| BEFORE
    C -->|"Restored"| AFTER

    style BEFORE fill:#ffcccc,stroke:#cc0000
    style AFTER fill:#ccffcc,stroke:#00cc00

Loading

_{Last reviewed commit: 8c5be4c}

[greptile-apps]

_{2 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[hcavarsan]

@yuneng-jiang this was a bug only on macos , ive made the PR #21899 with details on how i fixed issues