IGNITE-23303 Improve transaction performance.

[ascherbakoff]

https://issues.apache.org/jira/browse/IGNITE-23303

Major changes in PR:

1. Reduced number of clock ticks on txn path
2. Commit timestamp is propagated to client's observable timestamp for full txn to avoid unnecessary waiting.
3. Removed System.currentTimeMillis mocks
4. raft messages don't propagate HLC
5. Removed some contention points on txn path

Benchmark results (Xeon(R) Silver 4314):
32 threads
Before (rev d6d402d)
Benchmark (batch) (fsync) (partitionCount) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert 1 false 64 thrpt 20 442348.635 ± 14953.282 ops/s

After
Benchmark (batch) (fsync) (partitionCount) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert 1 false 64 thrpt 20 675504.465 ± 40299.612 ops/s

[vldpyatkov]

I would recommend rewriting this method in this way (without TODO). It can skip CAS in cases where the letest clock value is the most one and does not notify a listener in case where the clock is chaged without jump up.
@rpuch Could you approve this approach with the listener?

public HybridTimestamp update(HybridTimestamp requestTime) {
        long requestTimeLong = requestTime.longValue();

        while (true) {
            long now = currentTime();

            // Read the latest time after accessing UTC time to reduce contention.
            long oldLatestTime = this.latestTime;

            if (oldLatestTime >= requestTimeLong && oldLatestTime >= now) {
                return hybridTimestamp(LATEST_TIME.incrementAndGet(this));
            }

            if (now > requestTimeLong) {
                if (LATEST_TIME.compareAndSet(this, oldLatestTime, now)) {
                    return hybridTimestamp(now);
                }
            } else {
                long newLatestTime = requestTimeLong + 1;

                if (LATEST_TIME.compareAndSet(this, oldLatestTime, newLatestTime)) {
                    notifyUpdateListeners(newLatestTime);

                    return hybridTimestamp(newLatestTime);
                }
            }
        }
    }

[ascherbakoff]

Let's do it in separate ticket.
I've created https://issues.apache.org/jira/browse/IGNITE-23707

[sanpwc]

You've left lot's of dead code. E.g. there's no longer need to provide HybridClock to Loza nor store it in NodeOptions. It's just an example. I guess there's more code to remove since we no longer need. Curious why it's not a dedicated PR?

[ascherbakoff]

The scope of additional changes is small, so I decided to include it.

You've left lot's of dead code

I believe this doesn't break anything.
Created a ticket for further cleaning up:
https://issues.apache.org/jira/browse/IGNITE-23742

[sanpwc]

Ticket description is empty. Anyone who is not in context won't understand what to do.

[ascherbakoff]

Added the description.

[sanpwc]

COUNTER

[ascherbakoff]

🆗

[sanpwc]

That doesn't look right. Every request should send currentTimestamp. Within proposal you will send same time with multiple inserts within same transaction.

[ascherbakoff]

This is intended change and doesn't break txn ordering guaranties.
At txn map phase we need to sync with remote clocks to calculate proper commit timestamp, so sending ts value doesn't matter.
hlc.current can used as well, but it involves more overhead.