Use spaces instead of parentheses for SQL sanitization.
This still solves the problem of negative numbers creating a line comment, but this avoids breaking edge cases such as
set foo to $1
where the substitution is taking place in a location where an arbitrary expression is not allowed.
Fix CVE-2024-27304
SQL injection can occur if an attacker can cause a single query or bind message to exceed 4 GB in size. An integer overflow in the calculated message size can cause the one large message to be sent as multiple messages under the attacker's control.
Thanks to Paul Gerste for reporting this issue.
- Fix behavior of CollectRows to return empty slice if Rows are empty (Felix)
- Fix simple protocol encoding of json.RawMessage
- Fix *Pipeline.getResults should close pipeline on error
- Fix panic in TryFindUnderlyingTypeScanPlan (David Kurman)
- Fix deallocation of invalidated cached statements in a transaction
- Handle invalid sslkey file
- Fix scan float4 into sql.Scanner
- Fix pgtype.Bits not making copy of data from read buffer. This would cause the data to be corrupted by future reads.
- Fix: prepared statement already exists
- Improve CopyFrom auto-conversion of text-ish values
- Add ltree type support (Florent Viel)
- Make some properties of Batch and QueuedQuery public (Pavlo Golub)
- Add AppendRows function (Edoardo Spadolini)
- Optimize convert UUID [16]byte to string (Kirill Malikov)
- Fix: LargeObject Read and Write of more than ~1GB at a time (Mitar)
- Allow NamedArgs to start with underscore
- pgproto3: Maximum message body length support (jeremy.spriet)
- Upgrade golang.org/x/crypto to v0.17.0
- Add snake_case support to RowToStructByName (Tikhon Fedulov)
- Fix: update description cache after exec prepare (James Hartig)
- Fix: pipeline checks if it is closed (James Hartig and Ryan Fowler)
- Fix: normalize timeout / context errors during TLS startup (Samuel Stauffer)
- Add OnPgError for easier centralized error handling (James Hartig)
- Add CopyFromFunc helper function. (robford)
- Add PgConn.Deallocate method that uses PostgreSQL protocol Close message.
- pgx uses new PgConn.Deallocate method. This allows deallocating statements to work in a failed transaction. This fixes a case where the prepared statement map could become invalid.
- Fix: Prefer driver.Valuer over json.Marshaler for json fields. (Jacopo)
- Fix: simple protocol SQL sanitizer previously panicked if an invalid $0 placeholder was used. This now returns an error instead. (maksymnevajdev)
- Add pgtype.Numeric.ScanScientific (Eshton Robateau)
- Add CollectExactlyOneRow. (Julien GOTTELAND)
- Add OpenDBFromPool to create *database/sql.DB from *pgxpool.Pool. (Lev Zakharov)
- Prepare can automatically choose statement name based on sql. This makes it easier to explicitly manage prepared statements.
- Statement cache now uses deterministic, stable statement names.
- database/sql prepared statement names are deterministically generated.
- Fix: SendBatch wasn't respecting context cancellation.
- Fix: Timeout error from pipeline is now normalized.
- Fix: database/sql encoding json.RawMessage to []byte.
- CancelRequest: Wait for the cancel request to be acknowledged by the server. This should improve PgBouncer compatibility. (Anton Levakin)
- stdlib: Use Ping instead of CheckConn in ResetSession
- Add json.Marshaler and json.Unmarshaler for Float4, Float8 (Kirill Mironov)
- Fix: QCharArrayOID was defined with the wrong OID (Christoph Engelbert)
- Fix: connect_timeout for sslmode=allow|prefer (smaher-edb)
- Fix: pgxpool: background health check cannot overflow pool
- Fix: Check for nil in defer when sending batch (recover properly from panic)
- Fix: json scan of non-string pointer to pointer
- Fix: zeronull.Timestamptz should use pgtype.Timestamptz
- Fix: NewConnsCount was not correctly counting connections created by Acquire directly. (James Hartig)
- RowTo(AddrOf)StructByPos ignores fields with "-" db tag
- Optimization: improve text format numeric parsing (horpto)
- Fix: RowScanner errors are fatal to Rows
- Fix: Enable failover efforts when pg_hba.conf disallows non-ssl connections (Brandon Kauffman)
- Hstore text codec internal improvements (Evan Jones)
- Fix: Stop timers for background reader when not in use. Fixes memory leak when closing connections (Adrian-Stefan Mares)
- Fix: Stop background reader as soon as possible.
- Add PgConn.SyncConn(). This combined with the above fix makes it safe to directly use the underlying net.Conn.
- Fix: concurrency bug with pgtypeDefaultMap and simple protocol (Lev Zakharov)
- Add TxOptions.BeginQuery to allow overriding the default BEGIN query
- Replace platform specific syscalls for non-blocking IO with more traditional goroutines and deadlines. This returns to the v4 approach with some additional improvements and fixes. This restores the ability to use a pgx.Conn over an ssh.Conn as well as other non-TCP or Unix socket connections. In addition, it is a significantly simpler implementation that is less likely to have cross platform issues.
- Optimization: The default type registrations are now shared among all connections. This saves about 100KB of memory per connection.
pgtype.Type
andpgtype.Codec
values are now required to be immutable after registration. This was already necessary in most cases but wasn't documented until now. (Lev Zakharov) - Fix: Ensure pgxpool.Pool.QueryRow.Scan releases connection on panic
- CancelRequest: don't try to read the reply (Nicola Murino)
- Fix: correctly handle bool type aliases (Wichert Akkerman)
- Fix: pgconn.CancelRequest: Fix unix sockets: don't use RemoteAddr()
- Fix: pgx.Conn memory leak with prepared statement caching (Evan Jones)
- Add BeforeClose to pgxpool.Pool (Evan Cordell)
- Fix: various hstore fixes and optimizations (Evan Jones)
- Fix: RowToStructByPos with embedded unexported struct
- Support different bool string representations (Lev Zakharov)
- Fix: error when using BatchResults.Exec on a select that returns an error after some rows.
- Fix: pipelineBatchResults.Exec() not returning error from ResultReader
- Fix: pipeline batch results not closing pipeline when error occurs while reading directly from results instead of using a callback.
- Fix: scanning a table type into a struct
- Fix: scan array of record to pointer to slice of struct
- Fix: handle null for json (Cemre Mengu)
- Batch Query callback is called even when there is an error
- Add RowTo(AddrOf)StructByNameLax (Audi P. Risa P)
- Fix: Support v4 and v5 stdlib in same program (Tomáš Procházka)
- Fix: sql.Scanner not being used in certain cases
- Add text format jsonpath support
- Fix: fake non-blocking read adaptive wait time
- Fix: json values work with sql.Scanner
- Fixed / improved error messages (Mark Chambers and Yevgeny Pats)
- Fix: support scan into single dimensional arrays
- Fix: MaxConnLifetimeJitter setting actually jitter (Ben Weintraub)
- Fix: driver.Value representation of bytea should be []byte not string
- Fix: better handling of unregistered OIDs
- CopyFrom can use query cache to avoid extra round trip to get OIDs (Alejandro Do Nascimento Mora)
- Fix: encode to json ignoring driver.Valuer
- Support sql.Scanner on renamed base type
- Fix: pgtype.Numeric text encoding of negative numbers (Mark Chambers)
- Fix: connect with multiple hostnames when one can't be resolved
- Upgrade puddle to remove dependency on uber/atomic and fix alignment issue on 32-bit platform
- Fix: scanning json column into **string
- Multiple reductions in memory allocations
- Fake non-blocking read adapts its max wait time
- Improve CopyFrom performance and reduce memory usage
- Fix: encode []any to array
- Fix: LoadType for composite with dropped attributes (Felix Röhrich)
- Support v4 and v5 stdlib in same program
- Fix: text format array decoding with string of "NULL"
- Prefer binary format for arrays
tracelog.TraceLog
implements the pgx.PrepareTracer interface. (Vitalii Solodilov)- Optimize creating begin transaction SQL string (Petr Evdokimov and ksco)
Conn.LoadType
supports range and multirange types (Vitalii Solodilov)- Fix scan
uint
anduint64
ScanNumeric
. This resolves a PostgreSQLnumeric
being incorrectly scanned intouint
anduint64
.
- Fix simple query sanitizer where query text contains a Unicode replacement character.
- Remove erroneous
name
argument fromDeallocateAll()
. Technically, this is a breaking change, but given that method was only added 5 days ago this change was accepted. (Bodo Kaiser)
- Update puddle to v2.1.2. This resolves a race condition and a deadlock in pgxpool.
QueryRewriter.RewriteQuery
now returns an error. Technically, this is a breaking change for any external implementers, but given the minimal likelihood that there are actually any external implementers this change was accepted.- Expose
GetSSLPassword
support to pgx. - Fix encode
ErrorResponse
unknown field handling. This would only affect pgproto3 being used directly as a proxy with a non-PostgreSQL server that included additional error fields. - Fix date text format encoding with 5 digit years.
- Fix date values passed to a
sql.Scanner
asstring
instead oftime.Time
. - DateCodec.DecodeValue can return
pgtype.InfinityModifier
instead ofstring
for infinite values. This now matches the behavior of the timestamp types. - Add domain type support to
Conn.LoadType()
. - Add
RowToStructByName
andRowToAddrOfStructByName
. (Pavlo Golub) - Add
Conn.DeallocateAll()
to clear all prepared statements including the statement cache. (Bodo Kaiser)
- Fix: CollectOneRow prefers PostgreSQL error over pgx.ErrorNoRows
- Fix: some reflect Kind checks to first check for nil
- Bump golang.org/x/text dependency to placate snyk
- Fix: RowToStructByPos on structs with multiple anonymous sub-structs (Baptiste Fontaine)
- Fix: Exec checks if tx is closed
- Fix
driver.Valuer
handling edge cases that could cause infinite loop or crash
- Fix date encoding in text format to always use 2 digits for month and day
- Prefer driver.Valuer over wrap plans when encoding
- Fix scan to pointer to pointer to renamed type
- Allow scanning NULL even if PG and Go types are incompatible
- Fix 32-bit atomic usage
- Add MarshalJSON for Float8 (yogipristiawan)
- Add
[
and]
to text encoding ofLseg
- Fix sqlScannerWrapper NULL handling
github.com/jackc/pgtype
, github.com/jackc/pgconn
, and github.com/jackc/pgproto3
are now included in the main
github.com/jackc/pgx
repository. Previously there was confusion as to where issues should be reported, additional
release work due to releasing multiple packages, and less clear changelogs.
CommandTag
is now an opaque type instead of directly exposing an underlying []byte
.
The return value ResultReader.Values()
is no longer safe to retain a reference to after a subsequent call to NextRow()
or Close()
.
Trace()
method adds low level message tracing similar to the PQtrace
function in libpq
.
pgconn now uses non-blocking IO. This is a significant internal restructuring, but it should not cause any visible changes on its own. However, it is important in implementing other new features.
CheckConn()
checks a connection's liveness by doing a non-blocking read. This can be used to detect database restarts or network interruptions without executing a query or a ping.
pgconn now supports pipeline mode.
*PgConn.ReceiveResults
removed. Use pipeline mode instead.
Timeout()
no longer considers context.Canceled
as a timeout error. context.DeadlineExceeded
still is considered a timeout error.
Connect
and ConnectConfig
have been renamed to New
and NewWithConfig
respectively. The LazyConnect
option has been removed. Pools always lazily connect.
The pgtype
package has been significantly changed.
Previously, types had a Status
field that could be Undefined
, Null
, or Present
. This has been changed to a
Valid
bool
field to harmonize with how database/sql
represents NULL
and to make the zero value useable.
Previously, a type that implemented driver.Valuer
would have the Value
method called even on a nil pointer. All nils
whether typed or untyped now represent NULL
.
Previously, the type system combined decoding and encoding values with the value types. e.g. Type Int8
both handled
encoding and decoding the PostgreSQL representation and acted as a value object. This caused some difficulties when
there was not an exact 1 to 1 relationship between the Go types and the PostgreSQL types For example, scanning a
PostgreSQL binary numeric
into a Go float64
was awkward (see jackc/pgtype#147). This
concepts have been separated. A Codec
only has responsibility for encoding and decoding values. Value types are
generally defined by implementing an interface that a particular Codec
understands (e.g. PointScanner
and
PointValuer
for the PostgreSQL point
type).
All array types are now handled by ArrayCodec
instead of using code generation for each new array type. This also
means that less common array types such as point[]
are now supported. Array[T]
supports PostgreSQL multi-dimensional
arrays.
Composite types must be registered before use. CompositeFields
may still be used to construct and destruct composite
values, but any type may now implement CompositeIndexGetter
and CompositeIndexScanner
to be used as a composite.
Range types are now handled with types RangeCodec
and Range[T]
. This allows additional user defined range types to
easily be handled. Multirange types are handled similarly with MultirangeCodec
and Multirange[T]
.
LoadDataType
moved to *Conn
as LoadType
.
The Bytea
and GenericBinary
types have been replaced. Use the following instead:
[]byte
- For normal usage directly use[]byte
.DriverBytes
- Uses driver memory only available until next database method call. Avoids a copy and an allocation.PreallocBytes
- Uses preallocated byte slice to avoid an allocation.UndecodedBytes
- Avoids any decoding. Allows working with raw bytes.
pgtype
previously supported and was tested against lib/pq. While it will continue to work
in most cases this is no longer supported.
Previously, most Scan
implementations would convert []byte
to string
automatically to decode a text value. Now
only string
is handled. This is to allow the possibility of future binary support in database/sql
mode by
considering []byte
to be binary format and string
text format. This change should have no effect for any use with
pgx
. The previous behavior was only necessary for lib/pq
compatibility.
Added *Map.SQLScanner
to create a sql.Scanner
for types such as []int32
and Range[T]
that do not implement
sql.Scanner
directly.
Int2
, Int4
, Int8
, Float4
, Float8
, and Uint32
fields now include bit size. e.g. Int
is renamed to Int64
.
This matches the convention set by database/sql
. In addition, for comparable types like pgtype.Int8
and
sql.NullInt64
the structures are identical. This means they can be directly converted one to another.
- Extracted integrations with https://github.com/shopspring/decimal and https://github.com/gofrs/uuid to https://github.com/jackc/pgx-shopspring-decimal and https://github.com/jackc/pgx-gofrs-uuid respectively. This trims the pgx dependency tree.
Bit
andVarbit
are both replaced by theBits
type.CID
,OID
,OIDValue
, andXID
are replaced by theUint32
type.Hstore
is now defined asmap[string]*string
.JSON
andJSONB
types removed. Use[]byte
orstring
directly.QChar
type removed. Userune
orbyte
directly.Inet
andCidr
types removed. Usenetip.Addr
andnetip.Prefix
directly. These types are more memory efficient than the previousnet.IPNet
.Macaddr
type removed. Usenet.HardwareAddr
directly.- Renamed
pgtype.ConnInfo
topgtype.Map
. - Renamed
pgtype.DataType
topgtype.Type
. - Renamed
pgtype.None
topgtype.Finite
. RegisterType
now accepts a*Type
instead ofType
.- Assorted array helper methods and types made private.
- Removed
AcquireConn
andReleaseConn
as that functionality has been built in since Go 1.13.
Previously, the connection read buffer would allocate large chunks of memory and never reuse them. This allowed transferring ownership to anything such as scanned values without incurring an additional allocation and memory copy. However, this came at the cost of overall increased memory allocation size. But worse it was also possible to pin large chunks of memory by retaining a reference to a small value that originally came directly from the read buffer. Now ownership remains with the read buffer and anything needing to retain a value must make a copy.
Control over automatic prepared statement caching and simple protocol use are now combined into query execution mode.
See documentation for QueryExecMode
.
pgx now supports named arguments with the NamedArgs
type. This is implemented via the new QueryRewriter
interface which
allows arbitrary rewriting of query SQL and arguments.
The RowScanner
interface allows a single argument to Rows.Scan to scan the entire row.
CollectRows
andRowTo*
functions simplify collecting results into a slice.CollectOneRow
collects one row usingRowTo*
functions.ForEachRow
simplifies scanning each row and executing code using the scanned values.ForEachRow
replacesQueryFunc
.
Rather than every type that implemented Begin
or BeginTx
methods also needing to implement BeginFunc
and
BeginTxFunc
these methods have been converted to functions that take a db that implements Begin
or BeginTx
.
Previously, the code for building a batch went in one place before the call to SendBatch
, and the code for reading the
results went in one place after the call to SendBatch
. This could make it difficult to match up the query and the code
to handle the results. Now Queue
returns a QueuedQuery
which has methods Query
, QueryRow
, and Exec
which can
be used to register a callback function that will handle the result. Callback functions are called automatically when
BatchResults.Close
is called.
Previously, a batch with 10 unique parameterized statements executed 100 times would entail 11 network round trips. 1 for each prepare / describe and 1 for executing them all. Now pipeline mode is used to prepare / describe all statements in a single network round trip. So it would only take 2 round trips.
Internal logging support has been replaced with tracing hooks. This allows custom tracing integration with tools like OpenTelemetry. Package tracelog provides an adapter for pgx v4 loggers to act as a tracer.
All integrations with 3rd party loggers have been extracted to separate repositories. This trims the pgx dependency tree.