Skip to content

Bundlebase 0.8.0

Two themes drove most of what landed in 0.8: mutable data and making Bundlebase work better as an agent tool.

New Features

DELETE and UPDATE

Rows can now be deleted or updated. The updates and deletes happen as layers over the existing data rather than physically modifying it so the history stays intact.

# Delete rows
bundle.delete("price < 0")

# Update rows
bundle.update({"status": "'closed'"}, "last_login < '2024-01-01'")

Or SQL:

DELETE WHERE price < 0
UPDATE SET status = 'closed' WHERE last_login < '2024-01-01'

The big goal with the update and delete support is help you clean your datasets.

ALWAYS DELETE and ALWAYS UPDATE

Persistent mutation rules that apply automatically to every fetch — useful for filtering out bad data at the source rather than after every import.

bundle.always_delete("is_test = true")
bundle.always_update({"region": "'unknown'"}, "region IS NULL")

SHOW ALWAYS DELETES

Agentic Usage

A lot of small things in this release were aimed at making Bundlebase easier for an agent to explore and use without hand-holding.

The MCP server now supports multiple bundles in a single session, so an agent can work across datasets without restarting. list-bundles lets it discover what's available. SYNTAX <command> gives inline help on any command, and bundle_info.commands lists every supported SQL command — so an agent can figure out what's possible without needing external docs.

The HTTP connector means an agent can pull in data from a URL directly, with format auto-detected from content type. Combined with TEST CONNECTOR, it can validate a data source before committing anything.

The REPL's / meta-commands are replaced with proper SHOW commands that work the same way in both interactive and programmatic use — less special-casing needed.

You can now test a connector without importing it or creating a bundle first. Runs discover() then data() and returns the results as a stream.

MCP server now supports multiple bundles concurrently in one session.

Data Exploration

Whether you are driving bundlebase with an agent or own your own, I added new features to help explore and understand your data.

  • DESCRIBE DATA IN statement provides column profiling: min, max, null count, distinct count
  • PDF reports — embed live charts and tables in a markdown file, generate a PDF
  • SHOW COMMANDS — lists all supported SQL commands with bundle_info.commands
  • SYNTAX <command> — inline help for any command
  • list-bundles command

Additional Data Formats

TSV has been added as an attachable file format

Excel and standard JSON files can be returned by connectors, and bundlebase auto-converts them into a supported internal format.

Standalone CLI Binaries

Each release now ships binaries for macOS (arm64), Linux (x86_64), and Windows (x86_64) — no Python required. Download from the releases page and put on your PATH.

Performance

A few things landed that make a noticeable difference on larger datasets:

  • LRU block cache — decoded RecordBatch objects are cached between queries, giving around 13x speedup at 100K rows in benchmarks
  • Parquet filter pushdown — predicate pushdown is now enabled, so filters get pushed into the parquet reader rather than scanning everything
  • ZSTD compression — parquet files now use ZSTD instead of the default
  • Sorted-vec tombstones — DELETE filtering is significantly cheaper for high-delete-count bundles

See the performance guide for details on tuning.

pip install bundlebase==0.8.0