Contents
Changelog
All notable changes to ProvSQL are documented
in this file. It mirrors the release-notes section of the website
(provsql.org/releases) and is kept in
sync by the release.sh release-automation script.
[1.3.1] - 2026-05-04
A bug-fix release focused on repair_key / mulinput correctness,
plus a corrected upgrade path from 1.2.3 and documentation additions
(a fifth case study and expanded material in case studies 1 and 2).
No on-disk format change relative to 1.3.0; an
ALTER EXTENSION provsql UPDATE is enough.
Upgrade-script corrections
sql/upgrades/provsql--1.2.3--1.3.0.sqlshipped with 1.3.0 only carried the per-database mmap migration warning and missed two groups of SQL-surface changes that had landed inprovsql.common.sql/provsql.14.sqlduring the 1.3.0 dev cycle: the lazy-input-gate refactor ofadd_provenance/repair_key(commitf670b7f) and the schema-qualifiedprovsql.time_validity_viewreferences intimetravel,timeslice,history, andget_valid_time(commit1f59032). Users on 1.2.3 who ranALTER EXTENSION provsql UPDATE TO '1.3.0'ended up with a stale set of function bodies. The script in 1.3.1 has been corrected and now applies all the missing changes; users still on 1.2.3 reach a clean 1.3.0-equivalent SQL surface when they upgrade after 1.3.1.sql/upgrades/provsql--1.3.0--1.3.1.sqlapplies the same catch-up changes on the 1.3.0 → 1.3.1 path so that users already on 1.3.0 (who came through the broken upgrade) are brought back in sync. Fresh installs of 1.3.0 also run this script, but the CREATE OR REPLACE statements match the source already on disk, so it is a no-op for them.
Bug fixes
probability_evaluate(..., 'tree-decomposition')on circuits containingmulinputgates. Input gates produced byrepair_keycould share an internal id when their UUIDs were never materialised in the d-DNNF builder, causing the probability to be wrong and to vary from one session to the next on identical data. The aliasing has been removed (src/BooleanCircuit.cpp,src/dDNNFTreeDecompositionBuilder.cpp); a regression test (test/sql/treedec_mulinput.sql) covers the affected query shapes.Off-by-one in
BooleanCircuit::rewriteMultivaluedGates. The splitter that turns amulinputinto a chain of independent Bernoulli inputs produced non-deterministic probabilities under self-join +GROUP BYqueries. Fixed; the four built-in evaluation methods (default,'possible-worlds','tree-decomposition','monte-carlo') now agree onmulinput-bearing circuits.Shapley and Banzhaf computation on
mulinputcircuits.shapley(),shapley_all_vars(),banzhaf(), andbanzhaf_all_vars()previously walked throughmulinputgates and returned meaningless values. They now raise a clear error identifying the unsupported gate type.
Documentation
New Case Study 5: The Wildlife Photo Archive. A 30-photo / 13-species / 63-detection synthetic dataset demonstrates the
VALUESclause,repair_keyand themulinputgate (with the numerical effect of mutual exclusion made explicit viasr_boolexprandprobability_evaluate), probabilistic ranking versus naive confidence thresholding,EXCEPTwith monus, common table expressions, andexpected()aggregates. The case study is bundled (no external data download) and is part of the regression suite.Case Study 1 gains three steps in the circuit-inspection section: a tree-decomposition probability variant in the benchmark step, an
sr_boolexprstep on the Nairobi monus token, and a programmatic circuit-inspection step usingget_nb_gates,get_gate_type,get_children, andidentify_token.Case Study 2 gains two steps: a bulk Shapley/Banzhaf step using
shapley_all_vars/banzhaf_all_vars(contrasted with the per-variable cross-join from the existing Steps 13 and 14), and a step on arithmetic over aggregate results illustrating theagg_tokencast warning.Copy-to-clipboard buttons on every documentation code block (
sphinx-copybutton). A small JS shim (doc/source/_static/copybutton-shim.js) papers over an incompatibility betweensphinx-copybutton0.4.0 (the version in Ubuntu Noble’s apt) and Sphinx 9.
Infrastructure
- Release tarballs and CI workflows exclude the
studio/subdirectory for future developments.
Upgrade procedure
make install
In each database that uses ProvSQL:
ALTER EXTENSION provsql UPDATE;
The mmap circuit format is unchanged from 1.3.0; no migration is required.
[1.3.0] - 2026-05-04
Breaking change: per-database circuit storage
Prior to 1.3.0, the provenance circuit was stored in four flat files at
the root of the PostgreSQL data directory ($PGDATA/provsql_gates.mmap,
provsql_wires.mmap, provsql_mapping.mmap, provsql_extra.mmap),
shared across all databases in the cluster. Starting with 1.3.0, each
database gets its own isolated set of files under
$PGDATA/base/<db_oid>/.
Users upgrading from 1.2.x must migrate their circuit data before
upgrading. The new provsql_migrate_mmap tool handles this. If the
migration is skipped, existing circuit data becomes inaccessible (new
provenance queries still work, but provenance computed under the old
version is lost). The upgrade script detects old flat files and raises a
WARNING with recovery instructions if they are still present.
Upgrade procedure
Install the new ProvSQL binaries:
make installRun the migration tool as the
postgresuser:provsql_migrate_mmap -D $PGDATA -c <connstr>The tool reads the old flat files, collects root UUIDs from each database’s provenance-tracked tables, writes per-database files under$PGDATA/base/<db_oid>/, and deletes the old flat files on success.Restart PostgreSQL.
In each database that uses ProvSQL:
sql ALTER EXTENSION provsql UPDATE;
If you forgot step 2
If PostgreSQL has already been restarted with the new binaries before migrating, some empty per-database files may have been created. To recover:
- Delete the empty per-database files:
rm -f $PGDATA/base/*/provsql_*.mmap - Restart PostgreSQL.
- Immediately run
provsql_migrate_mmapbefore executing any provenance query.
Lazy input gate creation
add_provenance() no longer eagerly writes an input gate to the circuit
for every existing row in the table at the time it is called. Gates are
now created on first reference during a query, at the cost of a small
overhead on the first query that touches each row. This significantly
reduces the overhead of provisioning large tables.
Four case studies
Four worked examples have been added to the documentation and are included as regression tests:
- Case Study 1: The Intelligence Agency: simple introductory example with Boolean and why-provenance.
- Case Study 2: The Open Science Database: comprehensive example covering why-provenance, where-provenance, custom semirings, probabilities, Shapley and Banzhaf values.
- Case Study 3: Île-de-France Public Transit: Boolean provenance and formula inspection over GTFS transit data.
- Case Study 4: Government Ministers Over Time: temporal provenance
with
union_tstzintervalsand time-validity views.
Bug fixes
- Fix GROUP BY provenance aggregation silently dropped when ORDER BY referenced the semiring result column.
- Fix d-DNNF tree decomposition: deduplicate OR gate children to prevent double-counting in probability evaluation.
- Fix NULL dereference and out-of-bounds crashes in where-provenance on views.
- Fix temporal functions (
time_filter,time_range,in_interval) to use schema-qualifiedprovsql.time_validity_view, preventing failures whensearch_pathdoes not include theprovsqlschema. - Fix
sr_booleanevaluation when the provenance mapping uses integer values. - Fix where-provenance PROJECT gate positions for provenance tables that are not the first RTE in a query, causing empty locator sets on some PostgreSQL versions.
[1.2.3] - 2026-04-12
PGXN improvements
Prevent indexing of secondary documentation directories (
doc/source/,doc/tutorial/,doc/demo/,doc/aggregation/,doc/temporal_demo/,where_panel/) on the PGXN distribution page viano_indexinMETA.json.Document PGXN as an installation channel in the user guide, with a note that
pgxn installdoes not configureshared_preload_libraries.Add a GitHub Actions workflow (
.github/workflows/pgxn.yml) that automatically publishes releases to PGXN on version-tag pushes.
Documentation and repository housekeeping
Add
CODE_OF_CONDUCT.md.Add architecture dataflow diagram to the website overview page.
Replace
sudowith generic “as a user with write access to the PostgreSQL directories” wording across installation and contribution documentation.
[1.2.2] - 2026-04-11
In-place extension upgrades
ALTER EXTENSION provsql UPDATE is now supported, starting with this
release. A committed chain of upgrade scripts under sql/upgrades/
covers every previous release (1.0.0 → 1.1.0 → 1.2.0 → 1.2.1 → 1.2.2),
so users on any historical version can upgrade in place without
dropping and recreating the extension. The persistent provenance
circuit (memory-mapped files) is preserved across the upgrade: the
on-disk format has been binary-stable since 1.0.0, and the relevant
headers (src/MMappedCircuit.h, src/provsql_utils.h) now carry
explicit warnings so future contributors don’t break that guarantee
by accident.
A pg_regress regression test (test/sql/extension_upgrade.sql)
exercises the full chain end-to-end on every PostgreSQL version in
the CI matrix, installing the extension at 1.0.0 from a frozen
install-script fixture and walking it up to the current
default_version. See the new “Extension Upgrades” section of the
developer guide for the workflow contributors should follow when
making SQL changes.
Repository housekeeping and discoverability
CHANGELOG.mdat the repository root, mirroring the release notes published at provsql.org/releases. It is automatically kept in sync byrelease.sh.GitHub issue and pull-request templates under
.github/. The bug-report form prompts for PostgreSQL version, ProvSQL version, OS, a minimal SQL reproducer, and optional verbose-mode output; the PR template carries a contributor checklist and links to the developer guide.DockerHub image-version badge added to the README (
inriavalda/provsql) and a prose pointer on the website overview page.PGXN
META.jsonat the repository root, making ProvSQL ready for submission to the PostgreSQL Extension Network. Submission will happen once upstream approval lands; no change to the build or install flow in the meantime.CITATION.cffnow carries the Zenodo concept DOI (10.5281/zenodo.19512786) and a Software Heritage archive URL in itsidentifiersblock.
Infrastructure
release.shlearned to updateCITATION.cff,CHANGELOG.md, andMETA.jsonin sync withprovsql.common.controlandwebsite/_data/releases.yml, and to enforce the presence of an upgrade script (auto-generating a no-op when no SQL sources have changed since the previous tag).- CI workflows now fetch git tags so
git describeworks inside the pgxn-tools containers, which unblocks the Makefile’s dev-cycle upgrade-script generation. - The four build / docs workflows'
paths-ignorelists excludeMETA.json,.github/ISSUE_TEMPLATE/**, and.github/pull_request_template.md, so metadata-only edits do not trigger the full CI matrix any more.
No SQL-level changes
There are no changes to sql/provsql.common.sql or
sql/provsql.14.sql in this release. The SQL API, query
rewriter, semiring evaluators, and probability machinery are
unchanged from 1.2.1. The upgrade script 1.2.1 → 1.2.2 is
accordingly an empty placeholder.
[1.2.1] - 2026-04-11
Maintenance release headlining the new developer guide and laying the groundwork for long-term archival and citation.
Highlights
Developer guide (14 chapters, ~3500 lines): PostgreSQL extension primer, architecture, query rewriting pipeline, memory management, where-provenance, data-modification tracking, aggregation semantics, semiring and probability evaluation (including the block-independent database model and the expected Shapley / Banzhaf algorithm from Karmakar et al., PODS 2024), coding conventions, testing, debugging, and the build system. Cross-references to Lean 4 machine-checked proofs of the positive-fragment rewriting rules and the m-semiring axioms. See the new Developer Guide tab in the documentation.
User guide updates: expanded coverage of
expected(), thechooseaggregate, custom semiring evaluation, and diagnostic functions. “Formula semiring” has been renamed to “symbolic representation” throughout.CITATION.cff: standard citation metadata at the repo root. GitHub now shows a Cite this repository button that emits BibTeX and APA for the ICDE 2026 paper.Software Heritage archival is active — the full repository history is continuously preserved at archive.softwareheritage.org.
Zenodo integration enabled: starting with this release, every tagged version receives a persistent DOI (10.5281/zenodo.19512786).
Fixes
create_provenance_mapping_viewis now available on all supported PostgreSQL versions, not only PG 14+.External-tool tests (
c2d,d4,dsharp,minic2d,weightmc,view_circuit_multiple) now skip cleanly when the tool is not installed, instead of being removed from the test schedule.
Infrastructure
- Automated documentation coherence check runs in CI (validates every
:sqlfunc:/:cfunc:/:cfile:/:sqlfile:reference resolves to a live Doxygen anchor). - Mobile-friendly Doxygen and Sphinx output.
- CI speedups: concurrency groups, skip-on-tags, macOS
pg_isreadyrace fix. - In-place extension upgrades via
ALTER EXTENSION provsql UPDATEare supported starting with this release; upgrade scripts live undersql/upgrades/and the path is exercised by an automated CI test.
[1.2.0] - 2026-04-10
This release focuses on providing broader and more consistent support for SQL language features within provenance tracking. Systematic testing across a wide range of query patterns led to numerous bug fixes, new feature support, and clearer error messages for unsupported constructs.
New Features
CTE support: Non-recursive
WITHclauses now fully track provenance. Nested CTEs (CTE referencing another CTE) and CTEs insideUNION/EXCEPTbranches are supported. Recursive CTEs produce a clear error message.INSERT ... SELECTprovenance propagation: When both source and target tables are provenance-tracked,INSERT ... SELECTnow propagates source provenance to the inserted rows instead of assigning fresh tokens. A warning is emitted when the target table lacks aprovsqlcolumn.Correct arithmetic and expressions on aggregate results from subqueries: Explicit casts (
cnt::numeric), arithmetic (cnt + 1), window functions (SUM(cnt) OVER()), and expressions (COALESCE,GREATEST, etc.) on aggregate results from subqueries now produce correct values with a warning, using the original aggregate return type from the provenance circuit.UNION ALL with aggregate columns:
UNION ALLof queries returning aggregate results now works correctly.
Bug Fixes
Fixed crash when mixing
COUNT(DISTINCT ...)withprovenance()orsr_formula(provenance(), ...)in the same query.Fixed
COUNT(*)returning NULL instead of0 (*)on empty results withoutGROUP BY.Fixed
provenance_cmpfunction failing with “function uuid_ns_provsql() does not exist” whenprovsqlwas not insearch_path.
Improved Error Messages
provenance_evaluateon unsupported gate types now reports the specific gate type and suggests using compiled semirings.Subquery errors now read “Subqueries (EXISTS, IN, scalar subquery) not supported” instead of the misleading “Subqueries in WHERE clause”.
Clear error messages for unsupported operations on aggregate results:
DISTINCTon aggregates,UNION/EXCEPT(non-ALL) with aggregates,ORDER BY/GROUP BYon aggregate results from subqueries.Dropped redundant “by provsql” suffix from all error messages (the “ProvSQL:” prefix is already present).
Documentation
Updated supported/unsupported SQL features list with accurate coverage based on systematic testing.
Added documentation for
INSERT ... SELECTprovenance propagation.Expanded aggregation documentation with examples of casts, window functions,
COALESCE, andGREATESTon aggregate results.Added workaround guidance for unsupported features (use
LATERALfor correlated subqueries, explicit cast for comparison on aggregates).
[1.1.0] - 2026-04-09
Support for arithmetic on aggregate results
Queries performing arithmetic on aggregate results (e.g.,
SELECT COUNT(*)+1 or SUM(id)*10) are now supported. Previously,
these queries produced incorrect results because the planner hook
replaced aggregate references with agg_token values without
adjusting surrounding operator type expectations. This is handled by
adding implicit and assignment casts from agg_token to standard SQL
types (numeric, double precision, integer, bigint, text),
and by inserting appropriate type casts during query rewriting when
aggregate results are used inside operators or functions. A warning
is emitted when provenance information is lost during such
conversions.
Infrastructure improvements
- Versioned Docker image tagging (images are now tagged with the
release version in addition to
latest). - Improved release process: post-release version bump is now automated, and release tarballs exclude non-essential files (CI workflows, release script, branding, Docker, and website assets).
- CI fixes for macOS and documentation builds.
[1.0.0] - 2026-04-05
Initial official release of ProvSQL after 10 years of development. ProvSQL is now fully documented and usable in production.