0.12.4

Breaking Changes

• Cost ties always picks Flink from Flink and PostgreSQL, with cost-model set to DEFAULT (#2557)
  • Before: the order of an internal map picked the engine, so the plan could change with enabled-engines or a leftover build/ folder
  • Now: Flink on every compile
  • Affected: deployed plans with such a table in PostgreSQL; a recompile moves it to Flink, and the new Flink job cannot restore the state of the old one

Important

Before deploying a recompiled plan of a running pipeline:

1. Compare the Stage: line of each table in build/pipeline_explain.txt with the deployed build
2. For a table that changed from postgres to flink, either keep it in PostgreSQL with /*+ engine(postgres) */, or deploy the new plan as a new Flink job without the old job's state

See Optimizer Configuration.

Features

• Database tables and indexes in pipeline_explain for every table that a database stores (#2554)
  • Output: a --- physical --- part with one Table <name> (<engine>): section per database table, listing its primary key and selected indexes
  • Index lines: type and columns as in the index hint, such as index Orders_2_btree_c1c2 (btree): customerid, time; __pk_hash marked (generated hash key)
  • JSON: a database_tables array on each stored table in pipeline_explain.json
  • Changed format: the bare --- before each table's definition is now --- definition ---
  • Failed compiles: pipeline_explain.*, pipeline_source.sqrl and pipeline_mutation_database.json no longer written when a compile fails between DAG planning and the backward compatibility check
• New function typesafe_classify in stdlib.typesafe, from Flink SQL Runner 0.12.4: classifies each row of an insert-only table with TypeSafe AI's Jev model (DataSQRL/flink-sql-runner#411)
  • Setup: IMPORT stdlib.typesafe.*;
  • Output: input columns, one typed <question>_answer column per question, _model, and _error with on_error => 'NULL'
  • Question kinds: noul, choice, score
  • API key: pipeline.global-job-parameters.typesafe.api-key in engines.flink.config, or the TYPESAFE_API_KEY environment variable
  • Details: module README
• Agent skills on indexes and engine selection in implement-sqrl and configure-sqrl (#2554, #2557)
  • Index checks: the api-query-indexes guide reads a table's indexes from pipeline_explain.txt instead of deploy/plan/postgres.json
  • Engine selection: the cost-model values, the Flink pick on a tie, and the Stage: line that shows each table's engine

Fixes

• Misplaced compile errors in DISTINCT, IMPORT and EXPORT statements, hints, and tables defined in other files (#2543)
  • Before: a DISTINCT error pointed into the query that the statement is rewritten into; an IMPORT error pointed at the start of the statement
  • Now: errors point at the column, package path or hint name they are about
  • Other files: Flink errors in a table from a .table.sql file or an imported script point into that file
  • Empty definitions: fail with Missing query or expression after ':=', at the table name
  • Output: a % in a message is kept; no Java stack trace printed with a Flink error

Maintenance

• Flink SQL Runner 0.12.4, up from 0.12.3 (see its release notes) (7ba7d3614)
  • New function: typesafe_classify in the new stdlib-typesafe module
  • Late records after a restore: fixed in kafka-safe and upsert-kafka-safe with idle advancement on; needs Describe on the consumer group for fresh starts with 'scan.startup.mode' = 'group-offsets'
  • Watermark alignment: scan.watermark.alignment.* now applies with SOURCE_WATERMARK() in kafka-safe and upsert-kafka-safe
  • Zero idle timeout: 'scan.watermark.idle-timeout' = '0' with SOURCE_WATERMARK() disables idleness instead of failing

Dependency	From	To	PR
Maven Enforcer Plugin	3.6.3	3.7.0	#2563
JaCoCo	0.8.15	0.8.16	#2561
Node.js (MCP Inspector test image)	26.12.1-alpine	26.13.0-alpine	#2564

---

Full Changelog: 0.12.3...0.12.4