This release adds direct PostgreSQL type scanning through database/sql on Go 1.27, improves compatibility with libpq connection strings and PostgreSQL date/time values, and includes further decoder hardening. See Changes for connection-string and date/time behavior changes that may affect existing applications.

Features

• stdlib: support Go 1.27's driver.RowsColumnScanner, allowing PostgreSQL types such as arrays and ranges to be scanned directly into Go values without pgtype.Map.SQLScanner. Existing database/sql scalar conversions and sql.Scanner behavior are preserved. The minimum supported Go version remains 1.25.
• Add Rows.TypeMap to expose the type map used to decode rows, including rows created by RowsFromResultReader that have no underlying Conn. Custom implementations of Rows, including mocks, must add this method.
• pgconn: add Config.MaxProtocolMessageBodyLen to configure the maximum incoming protocol message body size (carter-ya)
• pgconn: add ErrReadOnlyConnection, ErrReadWriteConnection, ErrPrimaryConnection, and ErrStandbyConnection sentinel errors for target_session_attrs validation, allowing callers to use errors.Is (Adrian-Stefan Mares)
• pgxpool: accept pool_ping_timeout in connection strings to configure Config.PingTimeout. The default is zero; zero and negative durations mean no timeout (1991santhu)

Changes

• Name-based row-to-struct mapping now matches explicit db tags case-insensitively, with exact matches taking precedence so tags can still distinguish quoted column names that differ only by case (AlisinaDevelo)

• pgconn: resolve the OS user account only when no user is supplied by the connection string, environment, or service file, avoiding unnecessary account lookups and crashes in some restricted container environments. Home-directory defaults for password, service, and TLS files remain available independently of the account lookup. On Unix these now use $HOME rather than the OS account's home directory (Mohamed MAACHE)

• pgtype: date, timestamp and timestamptz text values are now parsed and written by a hand-written parser and encoder for PostgreSQL's ISO date/time format instead of time.Parse and time.Format. Go's layout language cannot express a variable-width year or the BC era, which is the root of the bugs below. The text scan path is roughly 2.5x faster for timestamp and timestamptz. Bug fixes:

  • timestamp and timestamptz no longer silently move February 29 of a BC leap year to March 1 when encoding. time.Date(-4712, 2, 29, ...) was written as 4713-03-01 BC and is now written as 4713-02-29 BC. This affected ordinary four-digit BC years, not only extended-range ones. date was never affected.
  • timestamp and timestamptz can now scan BC leap days. 4713-02-29 BC previously failed with day out of range. date could already scan them.
  • Years past 9999 can now be scanned. 10000-01-02 03:04:05 previously failed to parse, so timestamp and timestamptz values at the high end of PostgreSQL's range were unreadable over the simple protocol and in any other text-format result.
  • time.Time arguments in the simple protocol now encode BC dates correctly, using the same timestamp encoder.
  • Fractional seconds beyond microsecond precision are rounded the way the server rounds them (round half to even, carrying into the rest of the value) instead of being kept at full precision. PostgreSQL never sends more than six fractional digits, so this only affects values from other sources.

  Behavior changes:

  • date now rejects impossible dates instead of normalizing them. 2024-02-30 returned 2024-03-01 and 2024-13-01 returned 2025-01-01; both are now errors. timestamp and timestamptz already rejected them.
  • All three types now reject values outside PostgreSQL's range for that type, in the binary format as well as the text format. PostgreSQL never sends out-of-range dates, so this only affects corrupt or hand-built input; the range is checked in both formats so that whether a value is accepted does not depend on QueryExecMode. timestamptz also rejects time zone displacements outside PostgreSQL's signed 32-bit seconds range, while accepting the wider offsets emitted for POSIX time zones, such as +16.
  • timestamptz values scanned from the text format are now returned in time.Local, or in ScanLocation when it is set, matching what the binary format has always returned. Previously the text path kept whatever location time.Parse derived from the offset the server sent, so the same value scanned in the two formats could report a different Location() and Zone(). The instant is unchanged, but everything that renders the location changes with it: Timestamptz.MarshalJSON now writes the client's offset rather than the server's, so a value the server sent as +05:30 marshals as 2024-01-01T13:34:05-08:00 on a UTC-8 client instead of 2024-01-02T03:04:05+05:30, and DecodeDatabaseSQLValue hands database/sql a time.Time in that same location. Set the codec's ScanLocation to time.UTC to pin the location regardless of the client's zone.
  • Error messages from these paths have changed.

• pgconn: connection URIs (postgres://...) are now parsed by a new parser designed to exactly match libpq's URI parser behavior instead of net/url, making pgx accept and reject exactly the same URIs as libpq (verified by differential fuzzing against libpq itself). Most connection strings are unaffected. Edge-case behavior changes, all matching libpq:

  • + in query values is literal, no longer decoded as a space.
  • Malformed percent-encoding is a parse error instead of the parameter being silently dropped. %00 is rejected.
  • Leading/trailing spaces in URI components are trimmed; interior spaces are a parse error (encode them as %20).
  • # is ordinary data, not a fragment delimiter.
  • The userinfo terminator is the first @ before any / (previously the last @).
  • When a query parameter is repeated, the last occurrence wins (previously the first).
  • ssl=true is accepted as an alias for sslmode=require in URIs (JDBC compatibility). A repeated ssl key follows the same last-occurrence-wins rule as other repeated parameters, even across the rewrite to sslmode. If the final ssl value is not true, an independent explicit sslmode remains in effect.
  • Multiple hosts with mixed port specs are positionally aligned: postgres://h1,h2:5433/db now means h1:5432 and h2:5433 (previously both hosts got port 5433). A port list that is neither a single port nor exactly one port per host is an error (could not match N port numbers to M hosts), also for keyword/value connection strings.
  • An IPv6 address in a URI must be enclosed in brackets. A bare postgres://::1/db was previously accepted as host ::1; it is now read as an empty host followed by port :1 and fails with an invalid port error. Write it as postgres://[::1]/db.
  • Empty host list elements (e.g. h1,,h2) get the default host instead of being dropped. Likewise, an empty host in a keyword/value string (host=) now means the default host -- typically the Unix socket directory -- where it previously meant a TCP connection to an empty hostname.
  • An empty port (?port= in a URI or port= in a keyword/value string) now means the default port 5432 for the affected hosts; previously it was an invalid port error. Like any connection-string port, a present-but-empty port takes precedence over PGPORT.
  • ASCII control characters (tab, newline, ...) in a URI are ordinary data bytes, as they are to libpq; net/url rejected any URI containing one. The exception is a literal NUL byte, which is still rejected, as net/url did. (libpq never sees one -- C strings end at the first NUL -- but in Go a raw NUL could otherwise pass through into the NUL-delimited startup message and inject extra parameters.)

  Unlike libpq, unrecognized URI query parameters are still accepted (they become runtime parameters or pgx-specific options). Parse error messages avoid quoting the unredacted connection string and redact recognizable password fields on a best-effort basis. Invalid connection strings can be structurally ambiguous, so password redaction cannot be guaranteed for every malformed input.

• pgconn: keyword/value connection strings (host=... user=...) now match libpq's parser exactly, the same treatment the URI parser received above and verified the same way, by differential fuzzing against libpq itself. Most connection strings are unaffected. Behavior changes, all matching libpq:

  • A backslash escapes whatever character follows it and is dropped, where previously only \\ and \' were unescaped and every other backslash was kept. A value containing a backslash must now escape it, as libpq requires: sslcert=C:\path\to\cert reads as C:pathtocert and has to be written sslcert=C:\\path\\to\\cert. This mainly affects Windows certificate and key paths, which previously came through intact without doubling.
  • A trailing backslash in an unquoted value escapes the end of the string, so it is dropped and the value ends there; it was previously rejected with invalid backslash. Inside a quoted value the escaped terminator leaves the string unterminated, which is still an error.
  • Whitespace inside a keyword is an error (missing "=" after "us" in connection info string) instead of becoming part of the key. Whitespace around the = is unaffected. This most often shows up with an unquoted value containing a space: application_name=my app host=x previously set neither parameter and sent app host to the server as a runtime parameter, and now fails to parse.

  As with URIs, unrecognized keywords are still accepted where libpq rejects them, and an empty user= is still dropped so that PGUSER and the OS user still apply.

Fixes

• Keep the connection open after a recoverable PostgreSQL error from Begin or BeginTx (Victor Alejandro Sanz Ararat)
• Call TraceQueryEnd when Exec fails while deallocating invalidated cached statements (Chris Bandy)
• Deallocate a failed prepare using the statement name actually sent to the server, and skip cleanup if Parse never completed, avoiding leaked prepared statements and unnecessary cleanup errors (Eliran Ben-Zikri)
• Fix LoadTypes overwriting scalar codecs such as box and point with an incorrect ArrayCodec (Arsen Ozhetov)
• pgconn: retrieve field descriptions when cached descriptions are empty, such as for cursor FETCH statements, including batch and pipeline execution (water)
• pgconn: keep batch statement descriptions and result formats aligned when commands return no rows or when Batch.ExecStatement is mixed with other batch commands; preserve field descriptions for empty results
• pgconn: handle empty and comment-only queries in pipeline mode, discard stale statement data after bind errors, and return a nil result from Pipeline.GetResults on error
• pgconn: skip reading the password file when a password is already set (Jared Fowkes)
• pgxpool: treat non-positive MaxConnLifetime values as unlimited instead of immediately expiring connections (Aurelien Pillevesse)
• pgtype: support non-comma text array delimiters through ArrayCodec.Delimiter, including the semicolon delimiter used by box[]. LoadType and LoadTypes now load the delimiter from PostgreSQL (Sueun Cho)
• pgtype: quote text array elements containing internal whitespace (Louisa Huang)
• pgtype: quote and escape text range bounds containing delimiters, quotes, or backslashes, and distinguish empty string bounds from unbounded ranges (Sueun Cho)
• pgtype: preserve decimal precision in Numeric.ScanScientific and accept scientific notation in Numeric.UnmarshalJSON; reject out-of-range scientific exponents and preserve the original input in parse errors (Sueun Cho)
• pgtype: encode and decode numeric infinity in JSON as "Infinity" and "-Infinity" instead of encoding it as zero (Vladimir Saraikin)
• pgtype: treat a valid Numeric with a nil Int as zero in Int64Value, and return errors when converting NaN or infinity to an integer instead of panicking (Vladimir Saraikin)
• pgtype: fix an infinite loop when decoding binary numeric zero with a nonzero digit count (Vladimir Saraikin)
• pgtype: fix binary numeric digit-count overflow and trailing-byte handling. Binary encoding now rejects values whose digit count, weight, or scale cannot fit the wire format, while accepting the full unsigned digit-count range.
• pgtype: fix scanning through multiple pointer levels, including SQL NULL and XML values, and return an error instead of panicking when a pointer-to-pointer scan destination is nil (Rangel Reale)
• pgtype: use bounds-checked binary reads throughout the codecs and reject malformed lengths, counts, and trailing data. This includes fixes for panics on malformed records and truncated multiranges (Vladimir Saraikin), and validation of bit / varbit bit lengths against the actual data (g3m0sis).
• pgtype: return errors instead of panicking on malformed interval text (greymoth-jp), unterminated composite text fields, and text arrays whose dimensions and element counts disagree
• pgtype: cap the initial allocation estimate when parsing hstore text to avoid excessive allocation from unvalidated separator counts; valid hstores may still contain any number of pairs (AshSgDe29071999)
• pgtype: correct reversed bounds in integer scan error messages
• pgconn: a backslash as the last byte of a quoted value in a keyword/value connection string no longer panics with slice bounds out of range. host='a\ -- and the shorter ='\, reachable through pgx.ParseConfig and pgxpool.ParseConfig -- now return unterminated quoted string in connection info string, libpq's own message for the same input. The unquoted branch has been guarded since be69c1c1; the quoted branch carried the same unguarded increment since the parser was ported from pgx v3. Found by fuzzing (Maxim Korotkov)
• pgconn: error messages that embed the connection string now also redact password and sslpassword values supplied as URI query parameters; previously only the userinfo password was redacted. Redaction matches keys the way the parser does -- percent-encoded spellings such as pass%77ord= are recognized -- and masks the entire raw value, so a password containing a space cannot leak its tail into the error message. Credentials stranded outside the userinfo by a malformed URI are masked whole, and invalid-port errors no longer embed the offending text (which in a malformed URI can be a mislaid password). Redaction of invalid connection strings is necessarily best effort: their structure may be ambiguous, so some malformed inputs can still expose password text in an error.
• pgconn: ParseConfigOptions.ConnStringAllowedKeys no longer exempts an explicitly supplied empty port (?port= in a URI or port= in a keyword/value string) from the allow-list. Only the implied all-empty port list of a multi-host URI without ports (postgres://h1,h2/db) is exempt. An explicit empty port shadows PGPORT even though it is empty, so it must be allowed like any other user-supplied key. The URI-only ssl=true alias is accepted when either ssl or sslmode is allowed, and every ssl/sslmode spelling written in the URI is validated -- including occurrences superseded by later repeated parameters.
• pgconn: drain socket before close in asyncClose so context cancellation produces a TCP FIN instead of RST, avoiding "connection reset by peer" on the server / proxy (Sean Chittenden at CrowdStrike, Inc.)
• pgproto3: StartupMessage.Encode rejects a NUL byte in any parameter name or value instead of writing it. The startup message body is a run of NUL-delimited strings whose length is data-driven, so a NUL in a value ends that parameter and everything after it is read by the server as further parameters -- an application_name of x\x00user\x00admin changed the role the connection logged in as. libpq cannot reach this state because its parameters are NUL-terminated C strings. Connect now fails with nothing written to the wire, which covers settings that bypass connection string parsing: service files and direct assignment to Config.RuntimeParams, Config.User, or Config.Database.
• pgconn: keyword/value connection strings containing a NUL byte are rejected by ParseConfig, as URIs already were.