[PATCH 0/2] Rate limit certain warining messages

[Vladimir Davydov <vdavydov.dev@gmail.com>]

Certain warning messages can be printed quite often, e.g. "too long WAL
write" is printed by each fiber stalled on the WAL. If there are
hundreds of fibers and the IO subsystem becomes unresponsive from time
to time, the log may grow to the scale of gigabytes pretty quickly,
which complicates investigation of real problems.

Also, say() (the function used for logging) is quite CPU consuming.
E.g. in certain cases when readahead or net_msg_max is misconfigured,
"stopping input" messages may easily eat all available CPU time and
make the instance unresponsive.

To overcome those problems, this patch set introduces say_ratelimited()
helper, which is just like say() except it suppresses a messages if too
many messages of the same kind have already been printed in the last few
seconds. The new helper is used for all kinds of messages that may spam
the log.

https://github.com/tarantool/tarantool/issues/2218
https://github.com/tarantool/tarantool/commits/dv/gh-2218-rate-limit-certain-log-messages

Vladimir Davydov (2):
  Introduce basic rate limiter
  Rate limit certain warning messages

 src/box/iproto.cc          |  10 +++--
 src/box/txn.c              |   5 ++-
 src/box/vy_point_lookup.c  |   7 +--
 src/box/vy_quota.c         |   5 ++-
 src/box/vy_read_iterator.c |   9 ++--
 src/ratelimit.h            | 104 +++++++++++++++++++++++++++++++++++++++++++++
 src/say.h                  |  38 +++++++++++++++++
 test/unit/CMakeLists.txt   |   2 +
 test/unit/ratelimit.c      |  78 ++++++++++++++++++++++++++++++++++
 test/unit/ratelimit.result |  13 ++++++
 10 files changed, 256 insertions(+), 15 deletions(-)
 create mode 100644 src/ratelimit.h
 create mode 100644 test/unit/ratelimit.c
 create mode 100644 test/unit/ratelimit.result

-- 
2.11.0