Mensia是Erlang的OTP庫中一個帶有強事務的分佈式KV存儲引擎。能夠很是方便且高效的存儲Erlang的任何數據類型。而且該系統支持持久化和內存表混合使用,不強制要求全部節點的性質相同。
node
Mnesia的事務模型是2PC的模型,基本上能夠分爲如下幾個步驟git
讓全部參與的Node準備進行提交github
若是有一個Node回答No,讓全部回答yes的Node回滾緩存
若是全部Node都回答yes,讓全部Node提交服務器
mnesia_tm.erl中的t_commit完成了進行事務提交的準備工做,在這些準備工做中,arrange函數將緩存在ets中的數據操做轉化成prepare記錄和commit記錄。網絡
t_commit(Type) -> {_Mod, Tid, Ts} = get(mnesia_activity_state), %先把ETS表拿出來 Store = Ts#tidstore.store, if %單層事務 Ts#tidstore.level == 1 -> intercept_friends(Tid, Ts), %% N is number of updates case arrange(Tid, Store, Type) of {N, Prep} when N > 0 -> multi_commit(Prep#prep.protocol, majority_attr(Prep), Tid, Prep#prep.records, Store); {0, Prep} -> multi_commit(read_only, majority_attr(Prep), Tid, Prep#prep.records, Store) end; true -> %% nested commit Level = Ts#tidstore.level, [{OldMod,Obsolete} | Tail] = Ts#tidstore.up_stores, req({del_store, Tid, Store, Obsolete, false}), NewTs = Ts#tidstore{store = Store, up_stores = Tail, level = Level - 1}, NewTidTs = {OldMod, Tid, NewTs}, put(mnesia_activity_state, NewTidTs), do_commit_nested end.
而multi_commit函數完成事務的2PC部分,該函數分支是默認的Erlang事務使用的提交方式運維
%使用簡單的2PC進行, %1. 讓全部參與的Node準備進行提交 %2a.若是有一個Node回答No,讓全部回答yes的Node回滾 %2b.若是全部Node都回答yes,讓全部Node提交 multi_commit(sym_trans, _Maj = [], Tid, CR, Store) -> %% This lightweight commit protocol is used when all %% the involved tables are replicated symetrically. %% Their storage types must match on each node. %% %% 1 Ask the other involved nodes if they want to commit %% All involved nodes votes yes if they are up %% 2a Somebody has voted no %% Tell all yes voters to do_abort %% 2b Everybody has voted yes %% Tell everybody to do_commit. I.e. that they should %% prepare the commit, log the commit record and %% perform the updates. %% %% The outcome is kept 3 minutes in the transient decision table. %% %% Recovery: %% If somebody dies before the coordinator has %% broadcasted do_commit, the transaction is aborted. %% %% If a participant dies, the table load algorithm %% ensures that the contents of the involved tables %% are picked from another node. %% %% If the coordinator dies, each participants checks %% the outcome with all the others. If all are uncertain %% about the outcome, the transaction is aborted. If %% somebody knows the outcome the others will follow. %劃分全部的提交節點爲內存或磁盤 {DiscNs, RamNs} = commit_nodes(CR, [], []), %進入事務提交的準備狀態,這時候事務尚未真正的提交完成 Pending = mnesia_checkpoint:tm_enter_pending(Tid, DiscNs, RamNs), ?ets_insert(Store, Pending), %循環的發出提交申請到各參與的節點上 {WaitFor, Local} = ask_commit(sym_trans, Tid, CR, DiscNs, RamNs), %此處是死等,可是實際上也不是會完全死等 %什麼狀況會發生死等呢 %在ask_commit以後,對端節點死掉了,可是在下一次Erts心跳以前 %對端節點又啓動起來了,OK這就是個有意思的狀況 %全部節點都返回了贊成Outcome爲do_commit {Outcome, []} = rec_all(WaitFor, Tid, do_commit, []), ?eval_debug_fun({?MODULE, multi_commit_sym}, [{tid, Tid}, {outcome, Outcome}]), %向全部磁盤節點廣播提交 rpc:abcast(DiscNs -- [node()], ?MODULE, {Tid, Outcome}), %向全部內存節點廣播提交 rpc:abcast(RamNs -- [node()], ?MODULE, {Tid, Outcome}), case Outcome of do_commit -> mnesia_recover:note_decision(Tid, committed), do_dirty(Tid, Local), mnesia_locker:release_tid(Tid), ?MODULE ! {delete_transaction, Tid}; {do_abort, _Reason} -> mnesia_recover:note_decision(Tid, aborted) end, ?eval_debug_fun({?MODULE, multi_commit_sym, post}, [{tid, Tid}, {outcome, Outcome}]), Outcome;
腦裂
分佈式
傳說中的事務無限等待函數
腦裂的成因,主要是網絡不穩定,致使兩個節點長時間的失去聯繫,讓彼此都認爲對方掉線了。而這個時候,兩個節點都接收了大量的數據寫入。當兩個節點自動恢復集羣通訊的時候,沒法經過事務決議合併數據的時候纔會出現。
post
在ask_commit以後,對端節點死掉了,可是在下一次Erts心跳以前,對端節點又啓動起來了。OK,這就是這種有意思的狀況。基本上來說,這種事情發生的機率很是小,除非是設計失誤和對Erlang系統不熟悉濫用heart這東西產生的。
對於腦裂問題,沒有什麼特別想說的。首先讓運維作好內網通訊的管理,Mnesia集羣使用專用的內部交換機和交換機熱冗餘,一點都不過度。其次,作好腦裂發生的準備,在應用層面進行處理,能夠參考大神的https://github.com/uwiger/unsplit項目。
對於這個傳說中的問題,我本身在使用Mnesia的集羣中並無遇到過。解決這問題,首先,要搞清楚Erts集羣是怎麼互相探測是活的,能夠看到我前面的博文http://my.oschina.net/u/236698/blog/389737。其次,在整個集羣內部創建NTP服務器,保證整個集羣的對時穩定性。再次,使用heart時,不要嚴格按照那個心跳時間設置,至少要設置2.5倍節點之間心跳探測時間爲保活時間。