Skip to content
This repository was archived by the owner on Nov 21, 2026. It is now read-only.
This repository was archived by the owner on Nov 21, 2026. It is now read-only.

Served ArcadeDB bulk loads are refused by 26.10.1's 100 MiB HTTP body limit: raise it on the batch arms, disclose it, and make the client say why #237

Closed

Description

@tae898

The 26.10.1 campaign's served ArcadeDB graph arm failed at sf10 in both durability classes, repetition 1 (the stage then skipped the other repetitions): lean_http.ConnectionError: POST http://srv-...:2480/api/v1/batch/bench?wal=true&expectedEdgeCount=1938516&idMapping=true: [Errno 32] Broken pipe.

The server log of the failed cell says: Batch load on database 'bench' was refused after 65645 vertices and 915591 edges because the request body exceeded 'arcadedb.server.httpBodyContentMaxSize' (currently 104857600 bytes). Raise that setting or split the payload.

The setting is new in 26.10.1 (type Long, default 104857600 bytes = 100 MiB, -1 unlimited, measured on the wire). The October pin was a September commit that had no such limit, which is why October's served rows loaded.

l2_graph.py (_batch, build, build_messages) streams ONE POST per load. Persons plus KNOWS at sf10 is about 210 MiB, and the sf1full network (persons, KNOWS, and the message half) is several GiB, so every large served bulk load is refused. The e2 lane's served arm (e2_hybrid.py, ArcadeE2Server.build) loads through the same endpoint and is over the default at the 500k-product tier. The restart lane's arcadedb_graph_server loads through the l2 adapter.

Our lean_http.py client also hid the reason. A server that refuses early answers and closes the connection while the client is still sending, and the client raised BrokenPipe without reading the answer the server had already sent.

Decision (taken, not open): raise the limit to a value that never binds (-Darcadedb.server.httpBodyContentMaxSize=68719476736, 64 GiB) on the served arms that use the batch endpoint, keep the loader as one streaming request per load, and declare and disclose it end to end the way the arcadedb_query_cap override is: one constant in runner.py, a registered override with a row field read back from the running server, a generated sentence under every table that shows an affected arm, a row in PROTOCOL.md section 7 and one in CAMPAIGN.md section 7. It changes no result and no performance knob: it is a transport capacity limit that the harness's data volume exceeds, and the embedded arms have no HTTP body. Splitting the payload (the handler supports it, with RID references across requests) is out of scope because it changes the loader and the answer path.

Also in this change: lean_http.py reads whatever response the server already sent when sending the request fails with BrokenPipeError or ConnectionResetError, and puts its status and the first 500 bytes of its body into the ConnectionError, so the next refusal says why. The retry rules do not change.

Tests: the override registry and its carriers, a test that every arm that reaches the batch endpoint carries the flag with one value above the largest planned body, and a unit test of the client against a fake server that answers 413 and closes while the client is still sending.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions