[Tarantool-patches] [RFC v7 0/2] relay: provide downstream lag information

[Cyrill Gorcunov via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

Guys, take a look once time permit. Previous version is here
https://lists.tarantool.org/tarantool-patches/20210506214132.533913-1-gorcunov@gmail.com/

Not for merging yet! I think instead of applier_wal_stat structure we might need some
_commonly shared_ statistics structure, probably bound to WAL code and all other
threads will update it in a lockless way, because we might need to collect more
detais on WAL processing in future. I though of something like

	enum {
		WAL_ICNT__APPLIER_TXN_START_TM,

		WAL_ICNT__MAX,
	};

	struct wal_stat {
		int64_t	icnt[WAL_ICNT__MAX];
	} wal_st[VCLOCK_MAX];

and introduce

	wal_st__icnt_read(unsigned id);
	wal_st__icnt_write(unsigned id);

then applier will simply push last timestamp to WAL_ICNT__APPLIER_TXN_START_TM
counter, and later when we need to send ACK we use wal_st__icnt_read() to
fetch it back. We won't need to allocate some dynamic memory for it but
rather use statically preallocated shared between threads. Not sure though.

v4 (by Vlad):
 - add a test case
 - add docbot request
 - dropped off xrow_encode_vclock_timed, we use opencoded assignment
   for tm value when send ack
 - struct awstat renamed to applier_wal_stat. Vlad I think this is
   better name than "applier_lag" because this is statistics on WAL,
   we simply track remote WAL propagation here, so more general name
   is better for grep sake and for future extensions
 - instead of passing applier structure we pass replica_id
 - the real keeper of this statistics comes into "replica" structure
   thus unbound of applier itself
 - for synchro entries we pass a pointer to the applier_wal_stat instead
   of using replica_id = 0 as a sign that we don't need to update statistics
   for initial and final join cases
 - to write and read statistics we provide wal_stat_update and wal_stat_ack
   helpers to cover the case where single ACK spans several transactions

v7:
 - reworked the idea, so we always send last applied transaction's timestamp
   inside applier's ACK message
 - fixed changelong
 - in replica structure opencode the applier_txn_start_tm member
 - drop multiple ifs in applier_apply_tx
 - drop if statement in apply_synchro_row

Vlad you pointed

 > 4. Lag is updated in the relay thread, therefore you can't
 > simply read it in TX thread like you do in the diff block
 > above.

actually I can read the relay's lag in box.info() output, if the
relay object is removed then it won't have RELAY_FOLLOW state
so we're safe. This is what you meant?

branch gorcunov/gh-5447-relay-lag-7-notest
issue https://github.com/tarantool/tarantool/issues/5447

Cyrill Gorcunov (2):
  applier: send transaction's first row WAL time in the applier_writer_f
  relay: provide information about downstream lag

 .../unreleased/gh-5447-downstream-lag.md      |  6 ++
 src/box/applier.cc                            | 74 ++++++++++++---
 src/box/applier.h                             | 14 +++
 src/box/lua/info.c                            |  3 +
 src/box/relay.cc                              | 51 ++++++++++
 src/box/relay.h                               |  6 ++
 src/box/replication.cc                        |  1 +
 src/box/replication.h                         |  5 +
 .../replication/gh-5447-downstream-lag.result | 93 +++++++++++++++++++
 .../gh-5447-downstream-lag.test.lua           | 41 ++++++++
 10 files changed, 279 insertions(+), 15 deletions(-)
 create mode 100644 changelogs/unreleased/gh-5447-downstream-lag.md
 create mode 100644 test/replication/gh-5447-downstream-lag.result
 create mode 100644 test/replication/gh-5447-downstream-lag.test.lua

-- 
2.31.1