3.7.0

Date: 2026-04-22 Tag: 3.7.0

Overview

Tarantool 3.x is the recommended release series. Users of Tarantool 2.11 are encouraged to update to the latest 3.x release.

This release introduces 11 improvements and resolves 29 bugs since 3.6.2.

Please consider the full list of user-visible changes below.

Compatibility

Tarantool 2.x and 3.x are compatible in the binary data layout, client-server protocol, and replication protocol. It means upgrade may be performed with zero downtime for read requests and the order-of-network-lag downtime for write requests.

Please follow the upgrade procedure to plan your update actions.

Users of Tarantool 2.x may be interested in the compat options that allow to imitate some 2.x behavior. This allows to perform application code update step-by-step, not all-at-once.

Functionality added or changed

App

• Exposed new tcp_server_create and tcp_server_loop handles to the socket module so a listening socket and accept loop can be driven from different threads. The existing socket.tcp_server behavior remains unchanged and is equivalent to a pair of tcp_server_create(), tcp_server_loop() calls.

Core

• Introduced experimental support for application threads. Like the main thread an application thread runs an event loop and has a Lua state with most built-in modules available, but it cannot access the database directly. Application threads are configured with the new threads configuration option. For calling other threads, including the main thread, the new Lua module experimental.threads has been added (gh-12206).
• Usage of custom collations is now deprecated. Support for custom collations is scheduled to be dropped in Tarantool 4.0. Newer Tarantool versions will not recover if there are custom collations in the database (gh-12209).
• Introduced box.iproto.export() for exposing a Lua function to IPROTO_CALL requests without polluting the global Lua namespace (gh-10036).

Replication

• Added the replication.linearizable_quorum option controlling the number of replicas a node must synchronize with before performing a linearizable transaction. The default value is N - Q + 1, where N stands for the count of replicas in the replica set and Q is the evaluated replication_synchro_quorum value. The same formula was used internally prior to this change (gh-11876).

Config

• Now it is allowed to assign and delete nested fields inside schema nodes of the any type using deep paths (gh-10204).
• Added config:wait_status(statuses, [timeout]) to wait until the instance reaches one of the specified configuration statuses and returns the actual status observed when the wait finishes (gh-12389).
• Added the env option to config:get() to explicitly control whether TT_* environment variables are taken into account (gh-9824).

Log

• With the verbose option enabled in http.client, libcurl debug logs are now routed to the Tarantool log instead of stderr (gh-12392).

Metrics

• Updated the metrics submodule to 1.7.0.

Changes in 1.7.0:

• graphite: Added an ability to send metrics to multiple servers. Backward compatibility with previous plugin versions is preserved. From now on, the init method assigns a unique name to the created fiber using the incoming graphite server opts (if provided). Added a new stop() method to stop all fibers started by the plugin (gh-540).

• Deleting a replica via box.space._cluster:delete() no longer leaves information about that replica in the metrics (it was only removed after a cluster restart) (gh-538).

Tools

• Introduced the misc.memprof.available and misc.sysprof.available flags to determine whether the corresponding profiler is available for the current Tarantool build (gh-12215).

Bugs fixed

Core

• Fixed a bug where .xlog.inprogress files were not automatically deleted during server startup if the wal_dir value was not the default (gh-12081).

Vinyl

• Fixed vinyl_memory overrun during recovery when the memory limit is lower than at the previous start (gh-12390).

Replication

• Fixed a bug where replication could get stuck upon receipt of a 'promote' entry from a newly elected leader. This was likely to happen when the receiver itself had been a leader not long ago and still had some transactions in the synchro queue not yet written to the journal (gh-12205).
• Fixed a bug where a replica on subscribe could get stuck in the sync state for the duration of replication.timeout. This issue was observable in box.info.replication[...].upstream and could lead to temporary inconveniences, such as the inability to run box.ctl.promote() (when election_mode was not off) (gh-12025).

LuaJIT

Backported patches from the vanilla LuaJIT trunk (gh-12134). The following issues were fixed as part of this activity:

• Added ffi.abi("dualnum").
• Fixed stack checks in varargs calls in the GC64 build.
• Fixed stack checks in pcall()/xpcall() in the GC64 build.
• Fixed the allocation limit for the no-JIT build.
• Fixed handling of OOM errors on stack resizing in coroutine.resume() and lua_checkstack().
• Fixed recording of loops with a -0 step value or NaN control values.
• Fixed error reporting when an error occurs during error handling.
• Fixed a dangling reference for FFI callbacks.
• Fixed BC_UNM for a -0 argument in the dual-number mode.
• Fixed narrowing of unary minus in the dual-number mode.
• Fixed recording of string.byte(), string.sub(), and string.find().
• Fixed missing type conversion for BC_FORI slots in the dual-number mode.
• Fixed various corner cases in VM events.
• Fixed constructor index resolution recording in the JIT compiler.
• Fixed a UBSan warning in unpack().
• Fixed snapshot purge while recording a function header.

Config

• Fixed a startup failure when a Tarantool 3.x instance configured via the declarative config joined a 2.x replica set, where the master had no instance or replica set names stored in its snapshot (gh-10426).
• Fixed a bug in the supervised bootstrap strategy when a replica could hang on startup while joining an already bootstrapped replica set (gh-11318).

Datetime

• Fixed an assertion failure on an ambiguous case where day of year (yday, which defines calendar month and month day implicitly) and calendar month (without a month day) were both defined in the date text. Now such case is detected and an error is thrown (gh-11347).
• Fixed tzoffset calculations for cases like new({timestamp=x, tz='Zone'}) (gh-12412).
• Fixed inconsistency between dates produced by new({tzoffset=x}) and d:set({tzoffset=x}) where d.tz ~= '' comes before set() (gh-7680, gh-12416).
• Now datetime.new() and datetime_object:set() check the timestamp value to be in the valid range (gh-12147).

Election

• Fixed a bug where a newly elected leader could become writable (box.info.ro is false) immediately after box.info.election.state == 'leader' but before it claimed the synchronous transactions queue (box.info.synchro.queue.owner is not this node). This could occur only when the synchronous queue was previously unclaimed, which typically happens on cluster startup (gh-12333).
• Fixed a bug where box.ctl.promote() returned before the instance became writable (gh-10040).
• Fixed a bug where a replica could crash or exhibit undefined behavior when the election leader sent this replica a synchronous replication control command (promotion or demotion), but the replica received it just before performing a cascading journal rollback. This could happen if the wal.queue_max_size setting was reached (gh-12557).