PostgreSQL agent changelog
Release notes
2.3.2.2
Released: November 13, 2026
-
GSSD-1487: On an entity keyed by nullable columns, a delete matches a NULL key with IS NULL and an insert of a row the target already holds is looked up first and handled by the insert conflict strategy, so deletes no longer leave the row in place and replayed inserts no longer write a second copy;
-
GSSD-1487: A snapshot page or a refused insert batch that repeats a key is upserted in waves, the last row of the key winning, instead of inserting both rows or failing the page at every retry with "duplicate key value violates unique constraint";
2.3.2.0
Released: November 10, 2026
-
Deleting a single entity drops its replication slot, instead of leaving it in
pg_replication_slotsholding WAL untilmax_replication_slotsran out; a failed teardown keeps the entity on the agent so the deletion can be retried; -
GSSD-1346: Target: bulk delete and old-values fill run once per NULL pattern of the key (249 s down to 24 ms for 5,000 keys), and the upsert splits rows into runs sharing the same NULL keys (112 s down to 19 ms for 1,000 rows), so nullable key columns stay index keys;
2.3.1.4
Released: November 2, 2026
-
A map or a list is written as JSON text on the binary COPY and the JDBC batch insert too, instead of a map reaching jsonb as its toString form or being sent as an hstore;
2.3.1.2
Released: October 28, 2026
-
CDC: consecutive commits already waiting in the stream are merged into one emission, raising single-row commits from about 0.5k to about 18.7k rows per second. Truncates and chunked transactions are never merged, and a commit is confirmed to the slot only after its rows were emitted;
2.3.1.1
Released: October 27, 2026
-
Target: an object field from a NoSQL source is written as its JSON text instead of
{name=Ada, age=36}, and an array that holds a map or a nested array is written as a JSON array instead of[Ljava.lang.Object;@…;
2.3.1.0
Released: October 25, 2026
-
CDC (wal2json) and snapshots read BC dates, years after 9999 and a
timetzof24:00:00, and the bulkCOPYwrites BC dates and years after 9999 correctly; -
Target, bulk operations: the snapshot of a bulk entity ends when its last call flushes. Before, change capture could keep running as a snapshot, waiting for a full bulk batch and writing
NULLover the columns a partial update did not carry; -
Target, bulk operations: the staging table is emptied in the transaction of the merge, so a failed cleanup no longer merges the same rows again on the next flush;
2.3.0.3
Released: October 21, 2026
-
Target, bulk operations: the staging table keeps the type of its columns. A
moneycolumn is declared without a length (PostgreSQL rejectedmoney(18, 5)and the pipeline could not be created), an unconstrainednumericordecimalstays unconstrained instead of becomingnumeric(18,5)(values were rounded or overflowed), abit(n)orbit varying(n)keeps its length, and an unboundedvarcharorcharis declared without a length instead ofvarchar(0)orvarchar(MAX); -
Target, bulk operations:
inet,cidr,macaddr,macaddr8,interval,money,hstore, range and multirange,bitandbit varying, geometric (point,lseg,line,circle,box,path,polygon),tsvector,tsquery,txid_snapshot,pg_snapshot,pg_lsnandtimetzcolumns are loaded with the textCOPYand anoidcolumn with its four bytes, instead of being written in a wrong binary format. Columns of the common types stay on the binaryCOPY; -
Target, bulk operations: a
timestampinfinityor-infinityis written through theCOPY. Before,-infinitywas stored as the year 171979 andinfinityfailed theCOPY; -
Target, batch writes: a
bitcolumn is cast to its ownbit(n)and abitorbit varyingNULLis written asNULL,pg_lsn,txid_snapshot,pg_snapshotandoidvalues are cast to their type, and a delete matchesuuid,inet,bitandintervalkeys with the cast of the key’s type; -
Dates and timestamps before 1582 are read by the snapshot, and written by the binary
COPYand by batch writes, exactly (0001-01-01arrived as0001-01-03); -
wal2json CDC: the offset of a
timetzvalue is kept (01:02:03+01arrived as00:02:03+00), and atimestamptzwhose offset has seconds (+00:49:56) or whose year is 10000 is read instead of failing; -
wal2json CDC:
NaN,Infinityand-Infinityofrealanddouble precisioncolumns, also in a key, are replicated instead of arriving asNULL. The stream starts with the wal2json optionnumeric-data-types-as-string, which needs wal2json 2.6 or later. With an older wal2json the stream starts without it, these values stayNULL, and the agent logs a warning once. AnumericNaNhas no decimal value: the change now fails with a message instead of becomingNULL; -
An update of a row whose primary key is an array is compared by the elements of the key, so it is no longer replicated as a delete and an insert of a different key;
-
CDC resumed from a saved position skips only the changes of the transaction whose chunk was already delivered. Before, it dropped every change below that position, so a change written by a transaction still open while a shorter one committed was lost when its transaction committed later: the row never reached the target;
-
The replication slot is confirmed only up to the last commit the target has stored. Before, the slot was confirmed as each transaction was emitted, and the driver’s automatic flush moved it further on every reopened stream, so stopping and starting an entity, or changing its configuration, lost the commits a buffering target (bulk operations, file targets) had not stored yet: PostgreSQL never streams a slot from before its confirmed position;
2.3.0.2
Released: October 19, 2026
-
hstorecolumns are discovered and read as text (mapped to the string type) instead of being rejected as unsupported, on snapshot and wal2json CDC,hstorearrays included; -
Target: an
hstorecolumn is written with the schema the extension is installed in ("ext"."hstore"), so writes no longer fail when that schema is outside the agent user’ssearch_path; -
An update that changes the primary key is now sent as a delete of the row with the old key followed by an insert of the new row, as the other source agents do. Before, the row with the old key stayed on the target next to the new one. Message targets such as Kafka, Pub/Sub and Solace now receive a delete and an insert for such an update instead of a single update;
2.3.0.1
Released: October 15, 2026
-
wal2json CDC: range and multirange columns are serialised, an empty
byteastays empty, and schema and table names are escaped when tables are added; -
Arrays of a type outside the search path (such as an enum in another schema) keep their schema-qualified type;
-
Bulk operations: tables with array columns save and load correctly, including 2-D arrays and
NULLelements; -
Bulk operations: an insert and a delete, or two updates, of a row keyed by binary or array values in one batch are merged instead of leaving the row on the target or failing with a duplicate key;
-
The agent hands its pooled database back when its pipeline is deleted or detached, when its initialization fails, and, with recursion protection, when it closes. Before, the next pipeline writing to the same database could fail in a loop with
NoClassDefFoundError;
2.3.0
Released: October 13, 2026
-
Agent compatibility upgrade to Gluesync 2.3;
-
numericvalues keep every digit through wal2json CDC instead of being rounded through a double (12345678901234567890.1234567890arrived as12345678901234567000); -
Discovery returns schemas and tables in alphabetical order, so Query Studio pages no longer skip or repeat objects;