Microsoft SQL Server Table Polling

This adapter polls appended rows from one ordinary SQL Server table. It does not use Debezium and does not require Change Data Capture (CDC).

Source requirements

Select a non-null TINYINT, SMALLINT, INT, BIGINT, or scale-zero DECIMAL/NUMERIC sequence column backed by a single-column unique constraint. The table must be append-only and sequence values must become visible in increasing order. Updates, deletes, views, joins, and late lower-sequence commits are not supported.

Startup behavior

When the adapter starts, it records the current maximum sequence in memory. It emits only rows with a greater sequence value. Rows that exist before this watermark, including rows inserted while the adapter was stopped, are not replayed. The startup watermark is the activation boundary; it is captured before the polling scheduler starts.

Delivery

Rows are read using ordered keyset pagination and emitted individually. The in-memory cursor advances only after every row in a batch has been handed to the collector. A failed handoff retries that batch while the adapter remains running. Restarting the adapter captures a fresh watermark and intentionally provides no replay guarantee.

Polling limits and schema changes

The batch size defaults to 500 and the maximum rows per poll defaults to 10,000. The administrator can set the minimum allowed interval with SP_MSSQL_POLLING_MIN_INTERVAL_SECONDS (default: 1). Startup fails if the configured interval is below this value.

The output contains every source column, including the sequence, and no synthetic timestamp. Schema changes during an active run pause emission until the original column order and JDBC types return. A restart captures the current schema as the new baseline. Sequence regression caused by truncation or reseeding is not detected in this version.

SQL temporal values are emitted as epoch milliseconds and binary values as Base64 strings, matching the MSSQL connector conventions. Each poll uses and closes one JDBC connection. Connection login and query execution have finite timeouts.