Skip to main content

Indexing

The Indexing tab controls how Mitzu samples your warehouse when it indexes event and dimension tables to build the event and property catalog. The settings apply to the next indexing run: adding a table or re-indexing one. They do not re-index anything by themselves.

image

Time range​

  • Lookback window: how far back the indexer looks. The window ends on the workspace end date set on the Insights tab. A longer window finds events and properties that occur rarely, but each indexing query reads more data. Default: 30 days.

Sample size​

  • Event tables sample size: how many rows Mitzu samples from event tables. In a table that holds many events (one with an event name column), it is a cap per event name: Mitzu keeps up to this many rows of each event. One exception: when an indexing query covers a single event name and Randomise read order is off, Trino, Athena, Databricks, ClickHouse and Fabric skip the cap and read all of that event's rows in the lookback window. In a single-event table it is the number of rows sampled from the table. A larger sample finds more properties and filter values but makes indexing slower. Default: 2,000, maximum: 20,000.

  • Dimension tables sample size: how many rows Mitzu samples from each dimension table. Default: 20,000, maximum: 200,000.

Chunking for large tables​

By default Mitzu indexes a table with as few SQL queries as possible. On very large tables those queries can hit warehouse limits. Columns per chunk and Events per chunk split the work into smaller queries; Sampling steps reads more rows with extra queries.

  • Columns per chunk: for wide tables (many columns), index this many columns per query. Helps especially on data lakes with Parquet or ORC files. Leave empty to index all columns in one query. Empty by default.

  • Events per chunk: in tables that hold many events, index this many event names per query. Workspaces created since April 2026 start at 50, so chunking is on for them. Older workspaces show the field empty, which means every event is indexed in one query, until someone sets a value. Raise it to run fewer, larger queries, or leave it empty to index every event in one query.

  • Sampling steps: runs this many sampling queries per table instead of one. Each query reads a full sample (the sample size above) starting further into the table, so indexing reads up to this many times more rows and finds rarer values. It does not make any single query smaller. In tables that hold many events the extra queries currently read the same per-event sample again, so this setting only widens the sample of single-event and dimension tables. Leave empty for a single query. Accepts 1 to 20.

Read behaviour​

  • Randomise read order: reads rows in random order instead of the warehouse's default order, which can return a skewed slice of the table. The sample is more representative, but indexing is slower. Off by default.

Changes are saved automatically.