The vector clock on an object will increase in size for every new vnode_id which attem

Pruned clocks and failure of full-sync replication about riak_kv HOT 2 OPEN

martinsumner commented on August 16, 2024

Pruned clocks and failure of full-sync replication

from riak_kv.

Comments (2)

martinsumner commented on August 16, 2024

There is a workaround to "touch" (i.e. GET, then PUT back the object) all objects which have exceeded the prune size.

i.e.

{ok, CH} = riak:local_client().
TouchFun = fun({B, K, _C}) -> {ok, Obj} = riak_client:get(B, K, CH), ok = riak_client:put(Obj, CH) end.
QueriesFun = fun(Bucket) -> lists:map(fun(I) -> {fetch_clocks_range, Bucket, all, {segments, lists:seq((128 * I) + 1, 128 *  (I + 1)), small}, all} end, lists:seq(0, 511)) end.
TouchGreaterThanFun = fun(B, M) -> lists:foreach(fun(Q) -> lists:foreach(TouchFun,lists:filter(fun({_B0, _K0, C0}) -> length(C0) > M end, element(2, riak_client:aae_fold(Q, CH)))) end, QueriesFun(B)) end.
TouchGreaterThanFun(<<"clinicalsRecord">>, 50).

This is only possible when tictacaae is enabled, to allow for aae_folds.

from riak_kv.

martinsumner commented on August 16, 2024

It should be noted that read-repair will also override any sibling limit, so this can also lead to full-sync discrepancies. Hitting sibling limits is an indication of a broader fault, whereas pruning clocks is eventually inevitable, and so this is the primary concern.

from riak_kv.

Recommend Projects