Skip to content
HomePT · EN · ES · JA · ZH

When something fails

The pipeline finished successfully and the data lake is empty

Section titled “The pipeline finished successfully and the data lake is empty”

The most common symptom, and almost always configuration — not an error.

The pipeline does not fail for lack of configuration. With no database marked, the snapshot is exported, the Parquet files land in S3, and the copy job ignores them silently.

In order of likelihood:

Cause How to confirm
No database marked The console overview shows the alert
A new database appeared in the cluster Same, with the database name
Snapshot events turned off Same — in this case not even the export runs
Table filter excluded everything Under Bancos (Databases), see how many tables are marked

The copy job log also records what was ignored:

IGNORADO: 56 arquivos do banco "pagila" -- ele nao esta marcado
para ingestao, entao nada dele chega ao datalake.

If the table existed and stopped appearing, check under Partições (Partitions) whether there is a configuration for it with replace the direct copy on and the partition disabled. In that combination, the copy skips the table and the partitioner does not run — nobody writes it.

Turning the partition back on, or turning off “replace the direct copy”, fixes it.

Under Execuções (Runs), open the run to see the stages in order. The stage that failed carries the error message and the log.

The most common causes:

  • Memory during partitioning — a table with many distinct values in the partition column. See how many levels to use.
  • Cross-account permission — if step 3 of the cross-account installation was not done, the export fails when writing to the bucket.
  • Timeout — each job has one hour. A very large table with sorting configured is the most likely suspect.

A table that fails does not bring the others down: the pipeline tolerates isolated failures in partitioning and the crawler catalogs what was written.

If a partition_0 column shows up with values 1, 2, 3, the installation is on a version earlier than the one that flattens the RDS parts directory. Update the package and reprocess.

Two possible causes:

Table partitioned in Postgres. RDS exports the parent table and each child partition. Current versions detect and ignore the children; if you see tables like payment_p2024_01 next to payment, update the package.

Partition without “replace the direct copy”. By default, a partitioned table appears twice on purpose — the whole copy and the partitioned one. Turn the option on if you want only the partitioned one.

Cognito refuses callback URLs on http, except localhost. If the console serves over http, login fails before it starts. Run setup-https.sh.

If the browser warns that the certificate is not trusted, that is the self-signed certificate — it works, but to remove the warning use your own certificate through CERTIFICATE_ARN.

The load balancer is internal, on purpose: the console is not exposed to the internet. You reach it from inside the VPC, over VPN, or through a Session Manager tunnel.

Each job records its own version when it starts, and the run stores that. Under Execuções (Runs), the version of each component appears in the details — useful when behavior changed between updates.