Vladislav Shpilevoy

Lua `fiber:join()` is too unsafe

1

### Issue 1 One example: https://github.com/tarantool/tarantool/issues/7489. ### Issue 2 Another notably worse example below. * Create a file `test.c` with the following code: ```C #include "module.h" #include "lua.h" #include "lauxlib.h"...

```Lua -- -- Instance 1 -- -- Step 1 -- box.cfg{ listen = 3313, replication = {3313, 3314}, } box.schema.user.grant('guest', 'super') -- Step 3 -- s1 = box.schema.create_space('test1') _ =...

Our "system" fibers not only can be freely cancelled ATM (until #7473 is fixed), but also can be set as joinable and joined. Example: ```Lua net = require('net.box') box.cfg{listen =...

All triggers, visible to a user, crash on non-trivial changes

11

```Lua box.cfg{} s = box.schema.create_space('test') pk = s:create_index('pk') t1 = function() print('t1') end t2 = function() s:on_replace(nil, t1) s:on_replace(nil, t2) print('t2') end s:on_replace(t1) s:on_replace(t2) s:replace{1} ``` This leads to surprising...

``` Assertion failed: (test_stmt->txn->psn == 0), function memtx_tx_history_prepare_insert_stmt, file /Users/gerold/Work/Repositories/tarantool/src/box/memtx_tx.c, line 2189 ``` I got this on Tarantool 2.11.0-entrypoint-249-g373296fda. I don't have an isolated reproducer except a vshard branch: https://github.com/tarantool/vshard/tree/gerold103/mvcc-crash....

`net.box` worker fiber death leaves `"error_reconnect"` state but does not reconnect

1

```Lua net = require('net.box') box.cfg{listen = 3313} box.schema.user.grant('guest', 'super') c = net.connect(box.cfg.listen, {reconnect_after = 1}) c:ping() c._fiber:cancel() ``` ``` tarantool> c.state --- - error_reconnect ... ``` Reconnect doesn't happen anymore....

The option is useless after bootstrap. All it can do is break a working instance, after it is already booted. People constantly have problems with this option. The only purpose...

Retry join automatically (for a timeout)

1

Currently when join fails, the instance is terminated. Even if the error was not critical. For instance, the join-master appeared to be read-only. It looks like a config error, but...

`space:count()` with mvcc returns not the same as `space:select()`

1

Here is what I observe in one of vshard tests exclusively when mvcc is enabled: ``` 578: ilt.assert_equals(s:select{}, {{1, 1}, {2, 2}, {3, 1}, {4, 2}}) 579: ilt.assert_equals(s:count(), 4) ```...

Make rejoin cheaper

2

Our current rejoin algorithm - delete manually all the files and restart the node. It will download the entire dataset from scratch. That is quite expensive. And while normally it...