ishaangupta-yb/slack-agent/sizzle

Sizzle — Object Storage Analytics via DuckDB

Sizzle is Ishu's interface to storage statistics stored in DuckLake (Parquet or CSV files). It uses DuckDB's fast analytical SQL engine so you can ask about storage capacity, deduplication ratios, shard counts, and bandwidth directly from a Slack thread.

When to use

• You need to inspect storage metadata, shard distribution, or capacity trends.
• You want aggregated metrics (total deduplicated bytes, hot shards, bandwidth by day) without exporting CSVs.
• The data lives in DuckLake/Parquet/CSV files rather than Elasticsearch or MongoDB.

Tool: sizzle_query

Parameters:

• query (string, required): DuckDB SQL query to execute.
• files (string[], optional): File paths or globs relative to SIZZLE_DATA_DIR to query. Examples: ['2026/07/*.parquet', 'shards.csv']. Omit to run a pure SQL expression.
• format ("markdown" | "csv", default "markdown"): Output format. Markdown is best for Slack; CSV is best for copying into another tool.
• max_rows (number, default 50): Rows to return.

Authentication

Requires SIZZLE_DATA_DIR to point at the directory containing DuckLake files. No other credentials are needed for local files.

When querying remote DuckLake files hosted on an object store such as HuggingFace Hub, a local credential proxy can be enabled with HF_TOKEN, HF_PROXY_PORT, and HF_PROXY_TOKEN. The proxy only forwards GET requests under the configured dataset repository (set via HF_PROXY_REPO) and injects the token server-side so the runner never sees it.

Examples

Total storage by day for a Parquet dataset:

{
  "tool": "sizzle_query",
  "params": {
    "query": "SELECT date, SUM(bytes_deduplicated) AS dedup_bytes, COUNT(*) AS shards FROM __source_0 GROUP BY date ORDER BY date DESC",
    "files": ["2026/07/*.parquet"]
  }
}

Top 10 hottest shards by bandwidth:

{
  "tool": "sizzle_query",
  "params": {
    "query": "SELECT shard_id, SUM(bytes_out) AS total FROM __source_0 GROUP BY shard_id ORDER BY total DESC LIMIT 10",
    "files": ["bandwidth.parquet"]
  }
}

Capacity summary from a CSV export:

{
  "tool": "sizzle_query",
  "params": {
    "query": "SELECT SUM(raw_bytes) AS raw, SUM(dedup_bytes) AS dedup, ROUND(100.0 * (1 - SUM(dedup_bytes) / SUM(raw_bytes)), 2) AS savings_pct FROM __source_0",
    "files": ["capacity.csv"]
  }
}

Notes

• The tool wraps your query in a WITH CTE so it can apply max_rows. Do not add a trailing LIMIT unless you want to constrain the inner set before the row cap.
• DuckDB detects file type from extension. *.parquet uses read_parquet; *.csv uses read_csv_auto.
• If SIZZLE_DATA_DIR is not set, the tool returns a configuration message instead of failing.
• Large result sets are truncated to keep Slack messages readable.
• For sandbox safety, queries must be SELECT or WITH statements, cannot contain semicolons, comments, or destructive keywords, file paths cannot escape SIZZLE_DATA_DIR, and file-reading functions (for example read_parquet, read_text, read_json, parquet_scan, glob) are rejected inside the query. Always pass files via the files parameter.