Issue 1898: Implement isEnsembleAdheringToPlacementPolicy in RegionAwareEnsemblePlacementPolicy

[dragonls]

Master Issue: #1898

Motivation

Before this PR, isEnsembleAdheringToPlacementPolicy in RegionAwareEnsemblePlacementPolicy always return PlacementPolicyAdherence.MEETS_STRICT, which is not good to keep data highly available while using RegionAwareEnsemblePlacementPolicy , especially when one region is down.

Changes

Implement isEnsembleAdheringToPlacementPolicy in RegionAwareEnsemblePlacementPolicy.

The main implementation idea is that when all allocations are satisfied in different regions, it is considered PlacementPolicyAdherence.MEETS_STRICT, otherwise PlacementPolicyAdherence.FAIL.

[hangc0276]

@horizonzy Please help take a look at this PR, thanks.

[hangc0276]

@dragonls Thanks for your contribution. Would you please fix the failed check style? thanks.

[dragonls]

@dragonls Thanks for your contribution. Would you please fix the failed check style? thanks.

Fixed.

[horizonzy]

I have a question: why judge the regionsInQuorum.size() < 2?

[dragonls]

The purpose of doing this here is to align the implement in org.apache.bookkeeper.client.RegionAwareEnsemblePlacementPolicy#newEnsemble:

bookkeeper/bookkeeper-server/src/main/java/org/apache/bookkeeper/client/RegionAwareEnsemblePlacementPolicy.java

Lines 356 to 366 in 185fef6

	// Single region, fall back to RackAwareEnsemblePlacement
	if (numRegionsAvailable < 2) {
	RRTopologyAwareCoverageEnsemble ensemble = new RRTopologyAwareCoverageEnsemble(ensembleSize,
	writeQuorumSize, ackQuorumSize, REGIONID_DISTANCE_FROM_LEAVES,
	effectiveMinRegionsForDurability > 0 ? new HashSet<>(perRegionPlacement.keySet()) : null,
	effectiveMinRegionsForDurability, minNumRacksPerWriteQuorum);
	TopologyAwareEnsemblePlacementPolicy nextPolicy = perRegionPlacement.get(
	availableRegions.iterator().next());
	return nextPolicy.newEnsemble(ensembleSize, writeQuorumSize, writeQuorumSize,
	comprehensiveExclusionBookiesSet, ensemble, ensemble);
	}

When there is only one available region, fall back to RackAwareEnsemblePlacement.

[horizonzy]

It's a little weird.
Based on: writeQuorumSize = 3.

If regionsInQuorum < 2, the result may be MEETS_STRICT.

If regionsInQuorum = 2, the result is FAIL.

[dragonls]

Yes, a little weird.

Should we just remove the numRegionsAvailable < 2 judge in isEnsembleAdheringToPlacementPolicy? May I have your suggestion?

[horizonzy]

The PlacementPolicyAdherence enum has three values FAIL(1), MEETS_SOFT(3), MEETS_STRICT(5).
If there are enough different region, the result could be MEETS_STRICT. If there aren't enough different region, the each RackAwarePLacementPolicy MEETS_STRICT, the result could be MEETS_SOFT

[dragonls]

Updated.

[horizonzy]

Here, we can simplify the logic. We already know there is only one region.
Just need check the result of nextPolicy.newEnsemble.

1. If the result is PlacementPolicyAdherence.FAIL, return directly.
2. If the result is PlacementPolicyAdherence.MEETS_STRICT, then check whether minNumRacksPerWriteQuorum is more than 1. If it is, return PlacementPolicyAdherence.SOFT, otherwise return PlacementPolicyAdherence.MEETS_STRICT.