test: fix flaky feedback_daemon_metrics test

The patch fixes two mistakes that were made while writing the test.
Firstly, we should replace metrics module with an empty table before
every test case - otherwise, metrics can be collected before the test.
Secondly, all attempts to check if a required amount of metrics was
collected is pointless, even with margin - some environments are
incredibly slow. So let's check if some metrics were collected, and check
if there are not too many of them.

Also, the patch increases some timeouts to minimize the probability of
fail due to slow environment.

Follows up #8192
Closes tarantool/tarantool-qa#312

NO_CHANGELOG=test
NO_DOC=test

(cherry picked from commit d187b142)

test/box-luatest/gh_8192_feedback_daemon_metrics_test.lua

@@ -28,6 +28,8 @@ end)
 
 g.before_each(function()
     g.server:exec(function()
+        -- Replace module metrics with empty table to avoid metrics collection.
+        package.loaded["metrics"] = {}
         -- Daemon is reloaded when cfg option is updated.
         box.cfg{
             feedback_send_metrics = true,
@@ -54,24 +56,22 @@ g.test_metrics = function()
         end
         m._VERSION = "0.16.0"
         package.loaded["metrics"] = m
-        fiber.sleep(0.3)
+        fiber.sleep(0.5)
         local daemon = box.internal.feedback_daemon
         local fb = daemon.generate_feedback()
         -- This check may fail because of assertion failure in collect method.
         t.assert_type(fb.metrics, 'table')
-        -- Huge margin to prevent flakiness.
-        --
-        -- ASAN and GCOV builds looks very affected and macOS too.
-        --
-        -- In fact, this check verifies that *some* samples are collected. It
-        -- doesn't attempt to ensure correct timings, which is nearly
-        -- impossible in the general case, when many processes performs its own
-        -- workloads in parallel.
-        t.assert_almost_equals(#fb.metrics, 30, 25)
+        -- This check verifies that *some* samples are collected and that there
+        -- are not too many of them. It doesn't attempt to ensure correct
+        -- timings, which is nearly impossible in the general case, when many
+        -- processes performs its own workloads in parallel.
+        t.assert_gt(#fb.metrics, 0)
+        t.assert_lt(#fb.metrics, 60)
+
         fb = daemon.generate_feedback()
         -- Check is metrics were reset. Fiber yields on feedback generation, so
         -- metrics may be not empty.
-        t.assert(fb.metrics == nil or #fb.metrics < 5)
+        t.assert(fb.metrics == nil or #fb.metrics < 10)
     end)
 end
 
@@ -149,13 +149,13 @@ g.test_memory_limit = function()
         end
         m._VERSION = "0.16.0"
         package.loaded["metrics"] = m
-        fiber.sleep(0.2)
+        fiber.sleep(0.5)
         local daemon = box.internal.feedback_daemon
         local fb = daemon.generate_feedback()
         t.assert_type(fb.metrics, 'table')
         t.assert_equals(#fb.metrics, 5)
         box.cfg{feedback_metrics_limit = 1}
-        fiber.sleep(0.1)
+        fiber.sleep(0.3)
         fb = daemon.generate_feedback()
         t.assert_equals(fb.metrics, nil)
     end)