Agent Skills: oleander Iceberg Catalog

oleander Iceberg Catalog

Use this skill when reading from or writing to the oleander Iceberg catalog in a Spark job.

Catalog hierarchy

Tables are addressed as catalog.namespace.table:

• catalog: always oleander for the managed Lakekeeper-backed catalog
• namespace: a logical grouping (e.g., default, san_francisco, my_org)
• table: the table name

By default, default and telemetry namespaces are available out of the box. The telemetry namespace contains oleander-managed data you can query directly (for example oleander.telemetry.run_events, oleander.telemetry.traces, and oleander.telemetry.logs).

Examples:

oleander.default.sf_311
oleander.san_francisco.district_stats
oleander.my_namespace.results

Reading tables

Use spark.table() with the fully qualified name:

df = spark.table("oleander.default.sf_311")

Never construct table paths as raw S3 URIs. Always use the catalog name so Iceberg metadata and lineage are tracked correctly.

Writing tables

Append (add rows to an existing or new table):

df.writeTo("oleander.my_namespace.my_table").append()

Overwrite (replace table contents):

df.write.mode("overwrite").saveAsTable("oleander.my_namespace.my_table")

Use writeTo(...).append() for incremental pipelines. Use write.mode("overwrite").saveAsTable(...) when replacing full result sets each run.

Prefer Spark writes over driver writes

Avoid collecting data to the driver and then writing from Python memory. Keep writes as Spark DataFrame operations so Iceberg handles the transaction, partitioning, and metadata correctly.

Bad:

rows = df.collect()
# write rows from Python memory

Good:

df.write.mode("overwrite").saveAsTable("oleander.my_namespace.my_table")

Parameterize table names

Accept table names as arguments or environment variables so scripts are reusable:

import os, argparse

parser = argparse.ArgumentParser()
parser.add_argument("--input-table", default="oleander.default.sf_311")
parser.add_argument("--output-catalog", default="oleander.my_namespace")
args = parser.parse_args()

df = spark.table(args.input_table)
df.write.mode("overwrite").saveAsTable(f"{args.output_catalog}.results")

Namespace conventions

• Use lowercase, underscore-separated names for namespaces and tables.
• Keep the namespace tied to the domain or data source, not the job name.
• Avoid deeply nested namespaces; one level is usually sufficient.

Cache reused tables, then unpersist

If a table is read and used in multiple downstream transforms, cache it once and unpersist when done:

df = spark.table("oleander.default.sf_311")
df.cache()
# ... multiple transforms ...
df.unpersist()

Do not cache tables that are only used once.