Skip to content

request-log.md

Latest commit

 

History

History
77 lines (55 loc) · 3.12 KB

request-log.md

File metadata and controls

77 lines (55 loc) · 3.12 KB
Raw

Per-request serve log

Purpose

When two heavy requests overlapped, the serve engine left no trace, and clients hit the 30 s timeout without anyone seeing why. The per-request log records one line for each request, so overlap, queueing and slow commands can be found afterwards.

The log is off by default.

Enable

memnet serve --request-log PATH

or set the environment variable:

MEMNET_REQUEST_LOG=PATH

On the Pi the journal is RAM-only, so the journal does not persist. Use a file under the user's state directory. Create the directory first:

mkdir -p "$HOME/.local/state/memnet"
MEMNET_REQUEST_LOG=$HOME/.local/state/memnet/request.jsonl

Rotation

The file rotates by size.

  • MEMNET_REQUEST_LOG_MAX_BYTES: rotate when the file reaches this size. Default 10485760 (10 MiB).
  • MEMNET_REQUEST_LOG_BACKUPS: number of old files kept. Default 3 (minimum 1), giving PATH.1 to PATH.3.

Fields

Each line is one JSON object.

Field Meaning
received_ms Epoch milliseconds when the request frame was received.
finished_ms Epoch milliseconds when the reply was sent.
queue_wait_ms Time from the frame being fully read until the command starts. Includes drain-gate and session-lock waits (the work clock starts once the session lock is held).
server_ms Time from the command start until the reply is ready.
op Command words only, for example session open or mutate; a leading --version probe is recorded as --version. Never arguments.
session One-way label: s_ plus 8 hex digits of the session id. The engine has no session names, and the real mn_ id is never written.
bytes_in Bytes received in the request.
bytes_out Bytes sent in the reply.
lines_in Lines on stdin for the request.
lines_out Lines on stdout for the reply.
nodes_touched Currently always null. The dispatch layer does not know this.
outcome ok, an @ERR code (usage for an unknown option or command, internal for a crash), error, peer_gone, response_too_large or incomplete_frame.
conn_id Per-connection counter.
seq Request number on that connection. Always 1 today, because the protocol is one request per connection.

Notes

  • A line is written after the reply is sent.
  • Sink errors are swallowed. One warning is logged.
  • Bare connects, such as probes, are not logged.
  • No payload text, tokens or ids are written.
  • The log applies to TCP serve and to serve --ipc, which use the same handler.
  • memnet-mcp over HTTP reaches the engine through this serve when it uses TCP, so those requests are logged. In-process MCP mode does not go through serve and is not logged.

Sample line

{"received_ms":1793000000123,"finished_ms":1793000031456,"queue_wait_ms":1204.0,"server_ms":30120.0,"op":"mutate","session":"s_3f9a1c2e","bytes_in":18342,"bytes_out":211,"lines_in":96,"lines_out":3,"nodes_touched":null,"outcome":"ok","conn_id":57,"seq":1}

Finding overlaps

Requests that waited more than a second in the queue:

jq -c 'select(.queue_wait_ms>1000)' request.jsonl