umbra: the server crashes while loading a billion-row Parquet file with a large footer

Offen
  • Stimmt etwas nicht?
#3,292 0 Kommentare 0 Reaktionen 0 zugewiesene Personen Auf GitHub ansehen

Maintainer antworten meist innerhalb von 1 Tag

Dieses Issue hat noch niemand übernommen.

Bewertung

Schwierigkeit
4/5
Geschätzter Aufwand
3-5 Tage
Anfängerfreundlichkeit
25/100
Issue-Typ
Bug
Klarheit
Größtenteils klar
Aktivitätsstatus
Aktiv
Tech-Stack
shell, sql
Bereich
data, database

Rechercherichtung

The crash happens during the COPY in umbra/create.sql (around line 216) when umbra/load reads a Parquet file with about 2,360 row groups and a 31.6 MB footer. Start by rebuilding the reduced-row-group file with the clickhouse local command in the issue, then load it with umbra/load and capture the Umbra server log, which the report did not include. Done means the crash reproduces or is ruled out, with the footer size and row-group count recorded.

Vom Indexierungsmodell aus dem Issue-Text verfasst.

Beschreibung

The Umbra server crashed while loading a Parquet file of a billion rows: the client lost the connection during the COPY of umbra/create.sql:

CREATE TABLE
psql:create.sql:216: SSL SYSCALL error: EOF detected
psql:create.sql:216: error: connection to server was lost

This happened in the randomized benchmark (https://github.com/ClickHouse/ClickBench/pull/3271), which loads a generated dataset with the same schema as hits.parquet, a billion rows in one file, using umbra/install and umbra/load unchanged (image umbradb/umbra:26.10, machine c6a.metal with 384 GiB of memory). The file was written by ClickHouse 26.10 with its default row group size limit of 512 MB: more than a thousand row groups, and a footer of 35 MB. Hyper refused the same file because of its footer size (https://github.com/ClickHouse/ClickBench/issues/3291), so the size of the footer is a likely cause here too; the other systems of the benchmark loaded the same file. The log of the server was not collected.

Possible reproduction with the standard dataset: rewriting hits.parquet (100M rows, 226 row groups, a footer of 2.4 MB) with smaller row groups gives 2360 row groups and a footer of 31.6 MB:

clickhouse local --output_format_parquet_row_group_size 40000 --output_format_parquet_row_group_size_bytes 100000000000 \
    --query "SELECT * FROM file('hits.parquet') INTO OUTFILE 'hits-small-groups.parquet' TRUNCATE FORMAT Parquet"
clickhouse local --query "SELECT num_row_groups, metadata_size FROM file('hits-small-groups.parquet', ParquetMetadata)"
# 2360	31626081

then load it instead of hits.parquet with umbra/load.

The next run of the randomized benchmark writes about 250 row groups per file, as the original file has; its result will show whether Umbra loads a billion rows with a small footer, and I will add it here.

@toschmidt

🤖 Generated with Claude Code

Vorherrschende Sprache
Shell
Sterne
1.2k
Forks
334
Ø Merge
3 Std. 50 Min.
Gemergte PRs (30 T.)
710

Entwicklungsumgebung

  • Kein Dockerfile und keine Docker-Compose-Datei
  • Hat eine Pull-Request-Vorlage
  • Kein Beitragsleitfaden

Erste Schritte

  1. Lesen Sie das ganze Issue und danach den Beitragsleitfaden des Projekts.
  2. Schreiben Sie ins Issue, dass Sie es übernehmen — das erspart doppelte Arbeit.
  3. Forken Sie das Repository und arbeiten Sie in einem Branch.
  4. Öffnen Sie einen Pull Request, der die Issue-Nummer nennt.

Mehr aus ClickHouse/ClickBench

Alle Issues in ClickHouse/ClickBench

Ähnliche Issues

Weitere Issues zu Shell/Bash

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.