This is required to generate more observable railroad diagrams.

Антон Фетисов authored 1 month ago and Georgy Moshkin committed 1 month ago

Closes !915

It is the first step before porting Tarantool SQL tests into
Picodata. But we can also benefit right now and start writing less
verbose SQL feature tests.

Instead of linking against static vendored libcurl unconditionally, use
pkg_config/pkgconf to discover system libcurl and its version. At the
moment of writing, 8.4.0 is sufficient and thus can be used in
dynamic build mode.

Dmitry Rodionov authored 1 month ago and Dmitry Ivanov committed 1 month ago

It is needed because sometimes instance port is 5 symbols long, which
makes table wider by one symbol and test fails

Dmitry Rodionov authored 1 month ago and Dmitry Ivanov committed 1 month ago

The problem is that tarantool in 2.11.5 updated to quite recent curl
which is not available in repositories for our ci runner distro and
quite likely wont be available in distros we certify for. So for now
switch to static linking.

See https://git.picodata.io/core/picodata/-/issues/1299

The parameter with the old name still works, but the deprecation warning
is reported upon use. Will be removed a bit later.

Even though we're technically allowed to break backwards compatibility
I've done it this way, so that we have some examples of how to do such a
thing for future reference.

Simple subtraction overflow

Closes #1296

This is helpful when investigating incidents in case an instance of
picodata cannot start. Run `picodata tarantool -- -i` in the failing
instance's data directory, do `box.cfg { listen = 3301 }` in the
console. And then in another cluster's picodata instance you can do
`pico.raft_log({ remote = 'address:3301' })` to inspect raft log of that
failed cluster.

Remove the corresponding calls, add module-level doc-comments with
explanations on what the hell is going on.

This also fixes some flaky tests including #1294

Turns out blocking inside proc_replication_demote is not safe, as it may
lead to deadlocking if a replicaset master switchover happens during
a replicaset catch-up via raft snapshot. The symptom of the problem was
the flakiness of the tests (for example issue #1294).

Closes #1294

This reverts commit ec2b760a.

Антон Фетисов authored 1 month ago and Антон Фетисов committed 1 month ago

Use benchmark groups, calculate throughput for SQL operations, fix usage of `black_box` (outputs of bencher
function don't need to be black-boxed, since the framework already does it, but all inputs must be
properly black-boxed to avoid compiler optimizations).

Ideally we should add ability to configure the value, becase for tests
this timeout is unreasonably high.

Previously in our integration tests we would explicitly provide instance
names on startup. One of the reasons for that was, that we process all
the logging output from the instances by adding the prefix with the
instance name to them. After we stopped generating names explicitly this
feature broke, and the logs started being prepended with just the listen
port, which is not as useful.

This commit partially resolves this issue by making it so that logging
prefix is automatically updated after the instance is bootstrapped and
we find out it's name. This is implemented just by checking the
Instance.name member variable, which is updated when
Instance.instance_info() method is called. It is implicitly called in
Instance.wait_online() which happens in almost every test (often
implicitly as well). But if you notice that your test output doesn't
contain instance names at the start of the lines, you may want to call
Instance.instance_info() explicitly at some point.

Also note that there's an inherent problem with that approach -- any
log output which is handled before the instance name is known will be
prefixed with just the port number. Hopefully this is not a very big
problem in most cases...

A better solution would be to add the instance_name to all the logger
output automatically from picodata.
See #1303

Inside of string literals two subsequent single quotes
are parsed as a single one. The only exception is plugin values which
are considered to be json string and where single quotes are not escaped.

This reverts commit 6e91279f

Cartridge CI often fails with timeout errors on the vshard bootstrap
step. The reason is that cartridge uses default CALL_TIMEOUT_MIN in
vshard, that is 0.5 sec by default. As our CI is often busy with
concurrent jobs, current commit enlarges timeout 10 times up to
5 sec.

The main problem is that this test didn't wait properly until
the target instance is expelled. The correct logic could be
found in the test_join_expel_instance for audit. So, current
commit unifies the logic in both tests and adds some new methods
to the Cluster class.