分散合意パターン
分散システムの一貫性維持に不可欠な、分散合意アルゴリズムとパターンを理解し、実装します。
「分散合意パターン」はCoddyKit上の無料Erlang OTP: Distributed & Fault-Tolerant Systems Programmingレッスンです。 これはレッスン2/4です。 下記で完全なレッスンを無料で読むことができます。その後、ブラウザ内の組み込みコードエディタと24時間対応のAIチューターでハンズオン演習できます。 これはErlang OTP: Distributed & Fault-Tolerant Systems Programming学習パスの一部であり、ウェブとCoddyKitアプリ全体で進捗が同期されます。 Erlang OTP: Distributed & Fault-Tolerant Systems Programmingコースには全4レッスンが含まれています。
このレッスンの一部はまだ翻訳されておらず、英語で表示されています。
Agreeing in a Distributed World
Imagine multiple computers (nodes) needing to agree on a single outcome, even if some nodes fail or messages get lost. This challenge is called Distributed Consensus.
It's vital for maintaining data consistency and ensuring all parts of a system see the same "truth". Without it, your system might end up in a confused, inconsistent state.
The Hard Problem of Coordination
Achieving consensus is difficult because:
- Network Delays: Messages don't arrive instantly or in order.
- Node Failures: A computer might crash at any moment.
- Message Loss: Messages can be dropped by the network.
How do you ensure everyone agrees when communication is unreliable and participants can vanish?
Where Consensus Shines
Distributed consensus patterns are foundational for many critical system features:
- Leader Election: Deciding which node is the primary coordinator.
- Atomic Commits: Ensuring a transaction either fully completes on all nodes or completely fails on all.
- State Machine Replication: Keeping identical copies of data or application state across multiple nodes.
CAP and Consensus Trade-offs
The CAP Theorem states that a distributed system can only guarantee two out of three properties: Consistency, Availability, or Partition Tolerance.
Consensus algorithms typically prioritize Consistency and Partition Tolerance. This means during a network partition, the system might become unavailable for writes to prevent inconsistencies.
A Simple Agreement Protocol: 2PC
The Two-Phase Commit (2PC) protocol is a basic way to achieve atomic transactions across distributed nodes. It's often used in databases.
While not fully fault-tolerant (it can block if the coordinator fails), it's a great conceptual stepping stone to understanding more complex consensus algorithms.
The Coordinator: Orchestrating the Vote
In 2PC, one node acts as the Coordinator. Its job is to:
- Phase 1 (Prepare): Send a "prepare" or "vote request" message to all participating nodes.
- Phase 2 (Commit): Based on the votes, send a "commit" message if all voted "yes", or an "abort" message if any voted "no" (or timed out).
Participants: Deciding & Acting
Each Participant node in 2PC has these responsibilities:
- Phase 1 (Vote): When receiving "prepare", perform necessary checks. If ready to commit, reply "yes" and lock resources. Otherwise, reply "no".
- Phase 2 (Act): When receiving "commit", finalize the transaction. If "abort", roll back any changes and unlock resources.
Erlang Coordinator: Voting Process
Let's simulate a basic 2PC coordinator in Erlang. It spawns participants, sends a message, and collects their replies. This example simplifies error handling for clarity.
Note: This isn't production-ready 2PC, just an illustration of the message flow.
-module(coordinator).
-behaviour(gen_server).
-export([start_link/0, init/1, handle_call/3, handle_cast/2, handle_info/2, terminate/2, code_change/3]).
-export([propose/2]).
start_link() ->
gen_server:start_link({local, ?MODULE}, ?MODULE, [], []).
init([]) ->
{ok, []}.
propose(CoordinatorPid, Value) ->
gen_server:call(CoordinatorPid, {propose, Value}).
handle_call({propose, Value}, _From, _State) ->
% In a real system, participants would be registered or known
Pids = [
spawn(fun participant:start/0),
spawn(fun participant:start/0)
],
io:format("Coordinator: Proposing ~p to participants: ~p~n", [Value, Pids]),
% Phase 1: Prepare
Responses = [rpc:call(Pid, participant, prepare, [Value]) || Pid <- Pids],
FinalDecision =
case lists:all(fun(ok) -> true; (_) -> false end, Responses) of
true -> commit;
false -> abort
end,
io:format("Coordinator: All participants voted, decision: ~p~n", [FinalDecision]),
% Phase 2: Commit/Abort
[rpc:call(Pid, participant, FinalDecision, []) || Pid <- Pids],
{reply, FinalDecision, _State}.
handle_cast(_Msg, State) -> {noreply, State}.
handle_info(_Info, State) -> {noreply, State}.
terminate(_Reason, _State) -> ok.
code_change(_OldVsn, State, _Extra) -> {ok, State}.
% To run this example:
% 1. Compile both coordinator.erl and participant.erl
% 2. Start Erlang shell: erl
% 3. coordinator:start_link().
% 4. coordinator:propose(whereis(coordinator), "My Transaction").
% You should see output from both coordinator and participants.Erlang Participant: Voting & Acting
Here's how a participant process might respond to the coordinator. It simulates a "vote" and then acts on the "commit" or "abort" instruction.
This participant always votes 'ok' in this simplified version, but in reality, it would check its own state.
-module(participant).
-behaviour(gen_server).
-export([start_link/0, start/0, init/1, handle_call/3, handle_cast/2, handle_info/2, terminate/2, code_change/3]).
-export([prepare/1, commit/0, abort/0]).
start_link() ->
gen_server:start_link(?MODULE, [], []).
start() -> % Used by coordinator to spawn
{ok, Pid} = start_link(),
Pid.
init([]) ->
io:format("Participant ~p: Started.~n", [self()]),
{ok, #{} % State could hold transaction details
}.
prepare(_Value) ->
% In a real system, participant would check resources, lock them etc.
% For simplicity, always vote 'ok' here.
io:format("Participant ~p: Received prepare, voting 'ok'.~n", [self()]),
ok.
commit() ->
io:format("Participant ~p: Received commit, finalizing transaction.~n", [self()]),
ok.
abort() ->
io:format("Participant ~p: Received abort, rolling back transaction.~n", [self()]),
ok.
handle_call(_Msg, _From, State) ->
{reply, ok, State}. % Placeholder for any calls
handle_cast(_Msg, State) -> {noreply, State}.
handle_info(_Info, State) -> {noreply, State}.
terminate(_Reason, _State) -> ok.
code_change(_OldVsn, State, _Extra) -> {ok, State}.The Pitfalls of 2PC
While illustrative, 2PC has significant drawbacks:
- Single Point of Failure: If the coordinator crashes during Phase 2, participants might be left waiting indefinitely, holding locked resources. This is known as the "blocking problem".
- Performance: It requires multiple rounds of communication, which can be slow in high-latency networks.
These limitations necessitate more robust, non-blocking consensus algorithms like Paxos or Raft for truly fault-tolerant systems.
Quick Check: Consensus Roles
In the Two-Phase Commit (2PC) protocol, what is the primary responsibility of a Participant node in Phase 1 (Prepare)?
Recap: Agreement is Key
We've explored Distributed Consensus, understanding its importance for consistency in distributed systems and the challenges it presents.
We looked at Two-Phase Commit (2PC) as a basic protocol, understanding the roles of the Coordinator and Participants, and its key limitations. Erlang's message passing is a great foundation for building these patterns, but true fault-tolerant consensus requires more advanced algorithms.
よくある質問
「分散合意パターン」レッスンは無料ですか？
はい。「分散合意パターン」の完全なテキストはこのウェブで無料で読めます。インタラクティブに演習し（組み込みコードエディタと24時間対応のAIチューター）、Erlang OTP: Distributed & Fault-Tolerant Systems Programmingコースの残りをアンロックするには、CoddyKit PROにアップグレードしてください。 Erlang OTP: Distributed & Fault-Tolerant Systems Programmingコースには全4レッスンが含まれています。
「分散合意パターン」で何を学びますか？
分散システムの一貫性維持に不可欠な、分散合意アルゴリズムとパターンを理解し、実装します。 ブラウザで直接実行するハンズオンコードでErlang OTP: Distributed & Fault-Tolerant Systems Programmingを演習し、24時間対応のAIチューターがレッスンを進める中での質問に答えます。
Erlang OTP: Distributed & Fault-Tolerant Systems Programmingを始めるのに経験は必要ですか？
事前経験は必要ありません。CoddyKitのErlang OTP: Distributed & Fault-Tolerant Systems Programmingは初級者から上級者向けに構成されているため、ここから始めるか最初から始めて、自分のペースで進むことができます。 これはレッスン2/4です。
「分散合意パターン」レッスンにはどのくらい時間がかかりますか？
ほとんどのCoddyKitレッスンは約5～10分かかります。各レッスンはコンパクトでインタラクティブなので、着実に進歩し、ウェブとアプリ全体で正確に前回の場所から再開できます。
このErlang OTP: Distributed & Fault-Tolerant Systems Programmingレッスンでコードを書いて実行できますか？
はい。すべてのErlang OTP: Distributed & Fault-Tolerant Systems Programmingレッスンに組み込みコードエディタが含まれているため、ブラウザでリアルコードを書いて実行し、即座のAIフィードバックを取得できます。ローカル設定は不要です。