Tarantool 2.10.0-beta1 | Tarantool
Документация на русском языке
поддерживается сообществом

Tarantool 2.10.0-beta1

Released on 2021-08-20.

2.10.0-beta1 is the beta version of the 2 release series. It introduces 28 new features and resolves 56 bugs since the 2.8 version. There can be bugs in less common areas. If you find any, feel free to report an issue on GitHub.

Notable changes are:

  • Preliminary support of ARM64 CPU (GNU/Linux).
  • Preliminary support of Apple M1 CPU (macOS).
  • Consistent SQL type system.
  • Lightning fast net.box module.
  • Interactive transactions in binary protocol.
  • Small tuples are now stored more efficiently.

Tarantool 2.x is backward compatible with Tarantool 1.10.x in the binary data layout, client-server protocol, and replication protocol.

Please upgrade using the box.schema.upgrade() procedure to unlock all the new features of the 2.x series.

Some changes are labeled as [Breaking change]. It means that the old behavior was considered error-prone and therefore changed to protect users from unintended mistakes. However, there is a small probability that someone can rely on the old behavior, and this label is to bring attention to the things that have been changed.

  • The UUID field type is now part of the SCALAR field type (gh-6042).

  • The UUID field type is now available in SQL. A new UUID can be generated using the new SQL built-in function UUID() (gh-5886).

  • [Breaking change] The timeout() method of the net.box connection, marked deprecated more than four years ago (in 1.7.4), has been dropped. It negatively affected the performance of hot net.box methods, such as call() and select() if they were called without a specified timeout (gh-6242).

  • Improved net.box performance by up to 70% by rewriting hot code paths in C (gh-6241).

  • Introduced compact tuples that allow saving 4 bytes per tuple in case of small userdata (gh-5385).

  • Now streams and interactive transactions over streams are implemented in iproto. Every stream is associated with its ID, which is unique within one connection. All requests with the same non-zero stream ID belong to the same stream. All requests in the stream are processed synchronously. The next request will not start executing until the previous one is completed. If a request’s stream ID is 0, it does not belong to any stream and is processed in the old way.

    In net.box, a stream is an object above the connection that has the same methods but allows executing requests sequentially. The ID is generated on the client side automatically. If a user writes his own connector and wants to use streams, they must transmit the stream_id over the iproto protocol.

    The primary purpose of streams is transactions via iproto. As each stream can start a transaction, several transactions can be multiplexed over one connection. There are multiple ways to begin, commit, and rollback a transaction. One can do that using the appropriate stream methods, call, eval, or execute with the SQL transaction syntax. Users can mix these methods. For example, one might start a transaction using stream:begin(), and commit it with stream:call('box.commit') or stream:execute('COMMIT'). If any request fails during the transaction, it will not affect the other requests in the transaction. If a disconnect occurs while there is an active transaction in the stream, that transaction will be rolled back if it hasn’t been committed before the connection failure.

  • Added the new memtx_allocator option to box.cfg{}. It allows selecting an appropriate allocator for memtx tuples if necessary. The possible values are system for the malloc allocator and small for the default small allocator.

    Implemented the system allocator based on malloc. The slab allocator, which is used for tuple allocation, has a particular disadvantage—it is prone to unresolvable fragmentation on specific workloads (size migration). In this case, the user should be able to choose another allocator. The system allocator is based on the malloc function but restricted by the same quota as the slab allocator. The system allocator does not alloc all the memory at the start. Instead, it allocates memory as needed, checking that the quota is not exceeded (gh-5419).

  • Introduced the box.info.replication[n].downstream.lag field to monitor the state of replication. This represents the lag between the moment when the main node writes a certain transaction to its WAL and the moment it receives an ack for this transaction from the replica (gh-5447).
  • Introduced on_election triggers. The triggers can be registered via the box.ctl.on_election() interface and run asynchronously each time box.info.election changes (gh-5819).

  • Introduced support for LJ_DUALNUM mode in luajit-gdb.py (gh-6224).
  • Introduced preliminary support of GNU/Linux ARM64 and MacOS M1. In the scope of this activity, the following issues have been resolved:
    • Introduced support for the full 64-bit range of lightuserdata values (gh-2712).
    • Fixed memory remapping issue when the page left 47-bit segments (gh-2712).
    • Fixed M1 architecture detection (gh-6065).
    • Fixed variadic arguments handling in FFI on M1 (gh-6066).
    • Fixed table.move misbehaviour in case of table reallocation (gh-6084).
    • Fixed Lua stack inconsistency when xpcall is called on ARM64 with the invalid second argument (gh-6093).
    • Fixed BC_USETS bytecode semantics for closed upvalues and gray strings.
    • Fixed side exit jump target patching with regard to the range values of a particular instruction (gh-6098).
    • Fixed the issue on ARM64 where the current Lua coroutine wasn’t restored on the exception path (gh-6189).

  • Introduced the new method table.equals. It compares two tables by value with respect to the __eq metamethod.

  • Introduced new hash types—xxhash32 and xxhash64—in the digest module. (gh-2003).

  • Introduced fiber_object:info() to get info from a fiber. Works as require(fiber).info() but only for one fiber.
  • Introduced fiber_object:csw() to get csw from a fiber. Also, now csw (Context SWitch) of the new fiber is always equal to zero. Previously, it could be greater than zero (gh-5799).
  • Changed fiber.info() to hide backtraces of idle fibers (gh-4235).

  • The log module now supports symbolic representation of log levels. Now it is possible to specify levels the same way as in the box.cfg{} call (gh-5882).

    For example, instead of

    require('log').cfg{level = 6}

    it is possible to use

    require('log').cfg{level = 'verbose'}

  • Descriptions of type mismatch error and inconsistent type error have become more informative (gh-6176).

  • Removed explicit cast from BOOLEAN to numeric types and vice versa (gh-4770).

    For example, CAST(FALSE AS INTEGER) was 0 in version 2.8. Now it causes an error.

  • Removed explicit cast from VARBINARY to numeric types and vice versa (gh-4772, gh-5852).

  • Fixed a bug where a string that is not NULL-terminated could not be cast to BOOLEAN, even if the conversion would be successful according to the rules.

  • Now a numeric value can be cast to another numeric type only if the cast is precise. In addition, a UUID value cannot be implicitly cast to STRING/VARBINARY. Also, a STRING/VARBINARY value cannot be implicitly cast to a UUID (gh-4470).

  • Now any number can be compared to any other number, and values of any scalar type can be compared to any other value of the same type. A value of a non-numeric scalar type cannot be compared with a value of any other scalar type (gh-4230).

  • Removed SQL built-in functions from the _func system space (gh-6106).

  • The function is now looked up first in SQL built-in functions and then in user-defined functions.

  • Fixed incorrect error message in case of misuse of the function setting the default value.

  • The TYPEOF() function with NULL as an argument now returns NULL if the type cannot be determined from context. (gh-5956). Also, TYPEOF(-NaN) will now return DOUBLE, and TYPEOF(map_column) will now return VARBINARY instead of map.

    For example, SELECT TYPEOF(NULL) was BOOLEAN in version 2.8. Now it is NULL.

  • Reworked the SCALAR and NUMBER types in SQL. Removed the implicit cast from SCALAR to any other scalar type. Also, removed the implicit cast from NUMBER values to any other numeric type. It means that arithmetic and bitwise operations and concatenation are no longer allowed for SCALAR and NUMBER values. In addition, any SCALAR value can now be compared with values of any other scalar type using the SCALAR rules (gh-6221).

  • The field type DECIMAL is now available in SQL. Added an implicit cast from INTEGER and DOUBLE to DECIMAL and vice versa. DECIMAL can participate in arithmetic operations and comparisons with other defined numeric types (gh-4415).

  • [Breaking change] The argument types of SQL built-in functions are now checked in most cases during parsing. In addition, the number of arguments is now always checked during parsing (gh-6105).

  • [Breaking change] for the NUMBER data type. Arithmetic (+ * - / % & | ~) operations and bit-shift operations (>> <<) are now illegal.

    For example, SELECT number_column + 1 was legal in version 2.8. Now it causes an error.

  • [Breaking change] for the SCALAR data type. Arithmetic (+ * - / % & | ~) operations and bit-shift operations (>> <<) are now illegal. Concatenation (||) operations are now illegal. Values in SCALAR columns now have data type SCALAR, not the value’s data type.

    For example, TYPEOF(CAST(1 AS SCALAR)) was INTEGER in version 2.8. Now it is SCALAR.

  • [Breaking change]: Arithmetic operators must now have numeric operands. String operands are illegal.

    For example, SELECT 1 + '1' was 2 in version 2.8. Now it causes an error.

  • [Breaking change] in operations on SCALAR columns. Since the type of a value no longer determines whether an operation is valid, comparisons and functions that require a specific type no longer work.

    For example, table T has only one row with a scalar column containing 'a'. UPPER(scalar_column) was 'A' in version 2.8. Now it causes an error.

  • [Breaking change] for the HEX() function. STRING arguments are no longer acceptable; only VARBINARY arguments are allowed.

    For example, HEX('a') was '41' in version 2.8. Now it causes an error.

  • [Breaking change] for the POSITION() function. VARBINARY arguments are no longer acceptable; only STRING arguments are allowed.

    For example, POSITION(X'41',X'41') was 1 in version 2.8. Now it causes an error.

  • Previously csw (Context SWitch) of new fiber could be greater than 0, now it is always 0 (gh-5799).

  • Set FORCE_CONFIG=false for luarocks config to allow loading project-side .rocks/config-5.1.lua.

  • Fedora 34 builds are now supported (gh-6074).
  • Fedora 28 and 29 builds are no longer supported.

  • [Breaking change] fiber.wakeup() in Lua and fiber_wakeup() in C became NOP on the currently running fiber. Previously they allowed ignoring the next yield or sleep, which resulted in unexpected erroneous wake-ups. Calling these functions right before fiber.create() in Lua or fiber_start() in C could lead to a crash (in debug build) or undefined behaviour (in release build) (gh-6043).

    There was a single use case for the previous behaviour: rescheduling in the same event loop iteration, which is not the same as fiber.sleep(0) in Lua and fiber_sleep(0) in C. It could be done in the following way:

    in C:


    and in Lua:


    To get the same effect in C, one can now use fiber_reschedule(). In Lua, it is now impossible to reschedule the current fiber directly in the same event loop iteration. One can reschedule self through a second fiber, but it is strongly discouraged:

    -- do not use this code
    local self = fiber.self()
    fiber.new(function() self:wakeup() end)
  • Fixed memory leak on box.on_commit() and box.on_rollback() (gh-6025).

  • fiber_join() now checks if the argument is a joinable fiber. The absence of this check could lead to unpredictable results. Note that the change affects the C level only; in the Lua interface, fiber:join() protection is already enabled.

  • Now Tarantool yields when it scans .xlog files for the latest applied vclock and finds the right place to start recovering from. It means that the instance becomes responsive right after the box.cfg call even if an empty .xlog was not created on the previous exit.

    This fix also prevents the relay from timing out when a freshly subscribed replica needs rows from the end of a relatively long (hundreds of MBs) .xlog file (gh-5979).

  • The counter in N rows processed log messages no longer resets on each newly recovered xlog.

  • Fixed wrong type specification when printing fiber state change. It could lead to negative fiber IDs in the logs (gh-5846).

    For example,

    main/-244760339/cartridge.failover.task I> Instance state changed

    instead of proper

    main/4050206957/cartridge.failover.task I> Instance state changed
  • Fiber IDs are now switched to monotonically increasing unsigned 8-byte integers, so there is no ID wrapping anymore. It allows detecting fiber precedence by ID (gh-5846).

  • Fixed a crash in JSON update on tuple/space, where the update included two or more operations that accessed fields in reversed order and these fields didn’t exist. Example: box.tuple.new({1}):update({{'=', 4, 4}, {'=', 3, 3}}) (gh-6069).

  • Fixed invalid results of the json module’s encode function when it was used from the Lua garbage collector. For example, this could happen in functions used as ffi.gc() (gh-6050).

  • Added a check for user input of the number of iproto threads: value must be greater than zero and less than or equal to 1000 (gh-6005).

  • Changing a listed address can no longer cause iproto threads to close the same socket several times.

  • Simultaneously updating a key in different transactions does not longer result in a MVCC crash (gh-6131).

  • Fixed a bug where memtx MVCC crashed during reading uncommitted DDL (gh-5515).

  • Fixed a bug where memtx MVCC crashed if an index was created in the transaction thread (gh-6137).

  • Fixed a MVCC segmentation fault that arose when updating the entire space concurrently (gh-5892).

  • Fixed a bug with failed assertion after a stress update of the same key (gh-6193).

  • Fixed a crash where box.snapshot could be called during an incomplete transaction (gh-6229).

  • Fixed console client connection failure in case of request timeout (gh-6249).

  • Added a missing broadcast to net.box.future:discard() so that now fibers waiting for a request result wake up when the request is discarded (gh-6250).

  • box.info.uuid, box.info.cluster.uuid, and tostring(decimal) with any decimal number in Lua could sometimes return garbage if there were __gc handlers in the user’s code (gh-6259).

  • Fixed an error message that appeared in a particular case during MVCC operation (gh-6247).

  • Fixed a repeatable read violation after delete (gh-6206).

  • Fixed a bug where the MVCC engine didn’t track the select{} hash (gh-6040).

  • Fixed a crash in MVCC after a drop of space with several indexes (gh-6274).

  • Fixed a bug where the GC could leave tuples in secondary indexes (gh-6234).

  • Disallow yields after DDL operations in MVCC mode. It fixes a crash that took place when several transactions referred to system spaces (gh-5998).

  • Fixed a bug in MVCC that happened on rollback after a DDL operation (gh-5998).

  • Fixed a bug where rollback resulted in unserializable behavior (gh-6325).

  • Fixed possible keys divergence during secondary index build, which might lead to missing tuples (gh-6045).
  • Fixed the race between Vinyl garbage collection and compaction that resulted in a broken vylog and recovery failure (gh-5436).

  • Fixed the use after free in the relay thread when using elections (gh-6031).
  • Fixed a possible crash when a synchronous transaction was followed by an asynchronous transaction right when its confirmation was being written (gh-6057).
  • Fixed an error where a replica, while attempting to subscribe to a foreign cluster with a different replicaset UUID, didn’t notice it is impossible and instead became stuck in an infinite retry loop printing a TOO_EARLY_SUBSCRIBE error (gh-6094).
  • Fixed an error where a replica, while attempting to join a cluster with exclusively read-only replicas available, just booted its own replicaset, instead of failing or retrying. Now it fails with an error about the other nodes being read-only so they can’t register the new replica (gh-5613).
  • Fixed error reporting associated with transactions received from remote instances via replication. Any error raised while such a transaction was being applied was always reported as Failed to write to disk regardless of what really happened. Now the correct error is shown. For example, Out of memory, or Transaction has been aborted by conflict, and so on (gh-6027).
  • Fixed replication occasionally stopping with ER_INVALID_MSGPACK when the replica is under high load (gh-4040).
  • Fixed a cluster sometimes being unable to bootstrap if it contains nodes with election_mode set to manual or voter (gh-6018).
  • Fixed a possible crash when box.ctl.promote() was called in a cluster with more than three instances. The crash happened in the debug build. In the release build, it could lead to undefined behaviour. It was likely to happen if a new node was added shortly before the promotion (gh-5430).
  • Fixed a rare error appearing when MVCC (box.cfg.memtx_use_mvcc_engine) was enabled and more than one replica joined the cluster. The join could fail with the error "ER_TUPLE_FOUND: Duplicate key exists in unique index 'primary' in space '_cluster'". The same could happen at the bootstrap of a cluster having more than three nodes (gh-5601).

  • Fixed a rare crash with leader election enabled (any mode except off), which could happen if a leader resigned from its role while another node was writing something elections-related to WAL. The crash was in the debug build, and in the release build it would lead to undefined behaviour (gh-6129).
  • Fixed an error where a new replica in a Raft cluster tried to join from a follower instead of a leader and failed with the error ER_READONLY (gh-6127).

  • Fixed optimization for single-char strings in the IR_BUFPUT assembly routine.
  • Fixed slots alignment in the lj-stack command output when LJ_GC64 is enabled (gh-5876).
  • Fixed dummy frame unwinding in the lj-stack command.
  • Fixed detection of inconsistent renames even in the presence of sunk values (gh-4252, gh-5049, gh-5118).
  • Fixed the VM register allocation order provided by LuaJIT frontend in case of BC_ISGE and BC_ISGT (gh-6227).

  • Fixed a bug where multibyte characters broke space:fselect() output.
  • When an error occurs during encoding call results, the auxiliary lightuserdata value is not removed from the main Lua coroutine stack. Before the fix, it led to undefined behaviour during the next usage of this Lua coroutine (gh-4617).
  • Fixed a Lua C API misuse when the error is raised during call results encoding in an unprotected coroutine and expected to be caught in a different, protected coroutine (gh-6248).

  • Fixed a possible crash in case trigger removes itself. Fixed a possible crash in case someone destroys a trigger when it yields (gh-6266).

  • User-defined functions can now return a VARBINARY result to SQL (gh-6024).
  • Fixed assert when a DOUBLE value greater than -1.0 and less than 0.0 is cast to INTEGER and UNSIGNED (gh-6225).
  • Removed spontaneous conversion from INTEGER to DOUBLE in a field of the NUMBER type (gh-5335).
  • All arithmetic operations can now accept numeric values only (gh-5756).
  • Now QUOTE() returns the argument if the argument is DOUBLE, which is the same behavior as with other numeric types. For types different from numeric, the function returns a STRING (gh-6239).
  • The TRIM() function now does not lose collation when executed with the keywords BOTH, LEADING, or TRAILING (gh-6299).

  • Fixed MVCC interaction with ephemeral spaces: TX manager now ignores them (gh-6095).
  • Fixed loss of tuples after a conflict exception (gh-6132).
  • Fixed a segfault during update/delete of the same tuple (gh-6021).