erlang节点间传输eterm前,erlang会对eterm进行编码,如果eterm本身比较大,编解码会对系统性能造成较大的影响。对于某些特定的应用,例如备份系统,是可以通过某些方法进行优化的。
我们有一套N+1备份系统,enode里,一个worker流程结束时,会向备份enode发送状态数据,我们简单的使用{replica_handler, rep_enode} ! Msg来实现enode间消息的传递。随着系统日益庞大以及用户量的增长,平均每个worker的状态数据达到了4M左右,状态数据编解码对系统性能造成了不小的影响。
通过系统一些调试工具我们发现再我们的系统内, 单个worker相邻的两次状态数据,大部分是相同的。因此我们引入了差量备份(delta replica)机制。核心思想比较简单(类似于视频编码的关键帧和参考帧)
- 状态备份时,首次必须发送完整的状态信息给备份节点
- 随后i次状态备份,仅发送与上一次状态信息的差量部分给备份节点
- 当i%10 == 0时,强制发送完整的状态信息
详见
https://github.com/dead911/erlang/wiki/erlang-delta_replica
差量备份核心代码开源如下,仅供娱乐~
https://github.com/dead911/erlang/tree/master/delta_replica
%%% -------------------------------------------------------------------------
%%% Delta spec util
%%%
%%% delta_spec_util:get_delta_spec(Old, New) -> DeltaSpec
%%% to cacluate the delta spec of two given tuples
%%% Old - {user_info, {timestamp,}, {tid, }, {loc, {,}, na}}
%%% New - {user_info, {timestamp,}, {tid, }, {loc, {,}, na}}
%%%
%%% When delta_spec_level = , the output should be
%%% DeltaSpec -[{'$1',
%%% [{'$2',[{'$2',}]},
%%% {'$4',[{'$2',[{'$2',}]}]}]
%%% }]
%%% DeltaSpec indicates the difference between New and Old, each difference
%%% is presented in format [{<pos>, <value>}].
%%% {'$2',[{'$2',}]} the first '$2' implies the the nd element
%%% of New is different from Old. In this case, it is the 'timestamp'.
%%% The following [{'$2',}] means the second element of timestamp
%%% of New is differnt from of Old and the latest value is
%%%
%%% merge_delta_spec(OldSpec, NewSpec) -> NewSpec
%%% to merge two specs
%%% Considering the following scneario, when value changes from
%%% {hello, ,,} -> {hello, ,,} -> {hello, ,,}
%%% Two specs will be generated,
%%% OldSpec = [{'$4',}], {hello, ,,} -> {hello, ,,},
%%% NewSpec = [{'$3',}], {hello, ,,} -> {hello, ,,}
%%% After merge, NewSpec = [{'$3',}, {'$4',}]
%%%
%%% apply_spec_to_data(Data, Spec) -> NewData
%%% to apply spec to data
%%% Data - {, [ ,], }
%%% Spec - [{'$1, {'$2',[{'$1',12}]},{'$3',5}]}]
%%% NewData - {1, [12,3], 5}
%%% -------------------------------------------------------------------------