Extending location awareness for general join/leave support

[martinsumner]

Background information:

• The original problems with the old algorithm;
• The fix to the claim/join algorithm;
• The PR to fix the leave algorithm;
• The original rack awareness PR.

When joining a node, the following algorithms are attempted:

1 - A basic attempt to satisfy wants (vnodes required by the joining node) by asking node-by-node which vnodes can be passed on without breaking target_n_val (the claim_v2 algortihm). 2 - If Step 1 is unsuccessful, then attempt to stripe the all vnodes across all nodes (the sequential_claim algorithm). 3 - If Step 2 creates tail violations (i.e. if 0 < RingSize rem NodeCount < TargetNVal), resolve through the solve_tail_violations algorithm.

When leaving a node, the following algorithms are attempted:

1 - A basic attempt to perform a simple_transfer (vnodes are passed in turn to nodes that would not break target_n_val). 2 - Use sequential_claim as in join. 3 - Use solve_tail_violations extension to sequential_claim as in join

Ideally, in both cases Step 1 should succeed - as Step 2 will inevitable lead to a full cluster reorganisation (and hence a large volume of transfers).

As part of https://github.com/basho/riak_core/pull/967 location awareness was added to the sequential_claim algorithm (Step 2).

This issue is to document an ongoing investigation to these three problems:

• Under what conditions does the sequential_claim algorithm (both with and without the need for the solve_tail_volationa algorithm provide a location safe cluster;
• Can the claim_v2 (Step 1 for joins) and simple_transfer (Step 1 for leave) algorithms be extended to be location aware;
• Can the claim_v2 and simple_transfer algorithms be extended to reduce the scenarios in which cluster changes fallback to sequential_claim.

[martinsumner]

The initial condition to be tested is, will sequential_claim and solve_tail_violations consistently work if:

• 32 <= RingSize <= 2048;
• 6 <= count(Nodes) <= 64;
• count(Nodes) > RingSize;
• if max(count(NodesPerLocation)) = M, then there must be at least target_n_val locations where count(NodesPerLocation) is M.

To explain the last condition, if there are L locations, where L > TargetNVal, then at least TargetNVal locations must have M nodes, and the remaining (L - TargetNVal) locations must have =< M nodes.

[martinsumner]

The hypothesis above is incorrect. Even with these pre-conditions there are still failures with sequential_claim to support target_n_val.

e.g. RS 128, or RS 256 with 10 nodes split evenly across 5 locations.