0Pricing
Node.js Backend Development Bootcamp · 课时

构建读取模型与投影

从事件流派生经过优化的查询端投影，并使其最终保持一致。

构建读取模型与投影 是 CoddyKit 上的免费 Node.js Backend Development Bootcamp 课时。 这是第 3 节课，共 4 节。 你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。 这是 Node.js Backend Development Bootcamp 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Node.js Backend Development Bootcamp 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

Why Read Models Exist

In an event-sourced system, the write side stores facts as an append-only stream of events. That stream is great for capturing history but terrible for queries like "show me all open orders sorted by total".

The read side solves this. A read model (also called a projection) is a denormalized, query-optimized view derived purely from events. This is the Q in CQRS: commands mutate the event stream, queries hit read models.

  • One event stream can feed many read models, each shaped for a specific query.
  • Read models are disposable — you can delete and rebuild them from events at any time.
  • They are eventually consistent with the write side, not transactionally consistent.

A Projection Is a Left Fold

At its core, a projection is just a reduce over the event stream: you start with an initial state and apply each event in order to produce the next state.

Conceptually: readModel = events.reduce(apply, initialState). The apply function is a pure switch over event types. Anything you can fold, you can project.

// Pure projection: fold the event stream into a read model
const events = [
  { type: 'OrderPlaced', orderId: 'o1', total: 50 },
  { type: 'OrderPlaced', orderId: 'o2', total: 20 },
  { type: 'OrderPaid', orderId: 'o1' },
  { type: 'OrderCancelled', orderId: 'o2' },
];

function apply(state, event) {
  const next = { ...state };
  switch (event.type) {
    case 'OrderPlaced':
      next[event.orderId] = { status: 'placed', total: event.total };
      break;
    case 'OrderPaid':
      if (next[event.orderId]) next[event.orderId].status = 'paid';
      break;
    case 'OrderCancelled':
      delete next[event.orderId];
      break;
  }
  return next;
}

const readModel = events.reduce(apply, {});
console.log(readModel);
// { o1: { status: 'paid', total: 50 } }

A Projector Class

A real projector is more than a single fold. It owns three things:

  • handlers — a map from event type to an update function.
  • state — the read model it maintains (in memory, in a table, in a cache).
  • position — how far it has consumed the stream (the checkpoint).

Events it does not care about are simply ignored. This keeps each projector focused on exactly the data its queries need.

class OrderSummaryProjection {
  constructor() {
    this.state = new Map();
    this.handlers = {
      OrderPlaced: (e) => this.state.set(e.orderId, {
        status: 'placed', total: e.total, customer: e.customer,
      }),
      OrderShipped: (e) => {
        const r = this.state.get(e.orderId);
        if (r) r.status = 'shipped';
      },
    };
  }

  when(event) {
    const handler = this.handlers[event.type];
    if (handler) handler(event); // ignore unknown event types
  }

  query(orderId) {
    return this.state.get(orderId) ?? null;
  }
}

const proj = new OrderSummaryProjection();
proj.when({ type: 'OrderPlaced', orderId: 'o1', total: 99, customer: 'Ada' });
proj.when({ type: 'OrderShipped', orderId: 'o1' });
proj.when({ type: 'Irrelevant', foo: 1 });
console.log(proj.query('o1'));
// { status: 'shipped', total: 99, customer: 'Ada' }

Persisting to a Read Store

In production the read model lives in a database tuned for queries — Postgres, MongoDB, Redis, Elasticsearch — whatever fits the query shape. Each event handler issues an idempotent upsert against that store.

Note the SQL below is a fragment that assumes an order_summary table already exists, so it is illustrative rather than standalone-runnable.

// Postgres read-model upsert inside a projector (using node-postgres)
async function onOrderPlaced(pool, event) {
  await pool.query(
    `INSERT INTO order_summary (order_id, status, total, customer)
     VALUES ($1, 'placed', $2, $3)
     ON CONFLICT (order_id) DO UPDATE
       SET status = EXCLUDED.status,
           total  = EXCLUDED.total,
           customer = EXCLUDED.customer`,
    [event.orderId, event.total, event.customer]
  );
}

async function onOrderShipped(pool, event) {
  await pool.query(
    `UPDATE order_summary SET status = 'shipped' WHERE order_id = $1`,
    [event.orderId]
  );
}

Subscribing to the Event Stream

A projector does not poll forever blindly — it subscribes to the event store and receives events in order as they are appended. The contract every event store gives you:

  • Events arrive in global commit order (or per-stream order).
  • Each event carries a monotonic position (global offset or sequence number).
  • The subscription can start from a given position — essential for resuming.

You drive the projector by feeding each received event to its when method, then advancing the checkpoint.

// A minimal in-memory event bus a projector can subscribe to
class EventStore {
  constructor() { this.log = []; this.subs = []; }
  append(event) {
    const stored = { ...event, position: this.log.length + 1 };
    this.log.push(stored);
    for (const cb of this.subs) cb(stored);
  }
  subscribeFrom(position, cb) {
    for (const e of this.log) if (e.position > position) cb(e); // catch up
    this.subs.push(cb); // then live
  }
}

const store = new EventStore();
store.append({ type: 'OrderPlaced', orderId: 'o1', total: 10 });
let count = 0;
store.subscribeFrom(0, (e) => { count++; });
store.append({ type: 'OrderPaid', orderId: 'o1' });
console.log('events seen:', count); // 2 (1 catch-up + 1 live)

Checkpoints: Remembering Your Position

If your service restarts, you must not reprocess the entire stream from the beginning (slow) nor skip events (data loss). The fix is a checkpoint: persist the position of the last successfully processed event.

On boot, the projector reads its checkpoint and resumes the subscription from that position. The golden rule:

  • Process the event and update the read model.
  • Then advance and persist the checkpoint.
  • Ideally write the read-model change and the checkpoint in the same transaction.
// Resuming from a stored checkpoint
async function startProjector(store, checkpointRepo, projection) {
  const last = await checkpointRepo.load(projection.name); // e.g. 42

  store.subscribeFrom(last, async (event) => {
    await projection.when(event);              // 1. update read model
    await checkpointRepo.save(
      projection.name, event.position          // 2. advance checkpoint
    );
  });
}

// checkpointRepo example backed by a 'projection_checkpoints' table:
// load:  SELECT position FROM projection_checkpoints WHERE name = $1
// save:  INSERT ... ON CONFLICT (name) DO UPDATE SET position = $2

Idempotency: Surviving Redelivery

Because the checkpoint is saved after processing, a crash between "update read model" and "save checkpoint" means the same event will be redelivered on restart. This is at-least-once delivery, and it is normal.

Therefore every handler must be idempotent — applying the same event twice yields the same state. Two reliable techniques:

  • Use upserts instead of blind inserts (no duplicate rows).
  • For non-idempotent ops (counters, sums), record the last applied position per row and skip events at or below it.
// Guarding a running total against redelivery using a per-row version
function applyRevenue(state, event) {
  const row = state[event.customer] ?? { revenue: 0, lastPos: 0 };
  if (event.position <= row.lastPos) {
    return state; // already applied — skip duplicate
  }
  row.revenue += event.amount;
  row.lastPos = event.position;
  return { ...state, [event.customer]: row };
}

let s = {};
const paid = { type: 'OrderPaid', customer: 'Ada', amount: 30, position: 5 };
s = applyRevenue(s, paid);
s = applyRevenue(s, paid); // redelivered, ignored
console.log(s.Ada.revenue); // 30, not 60

Eventual Consistency and the Read-Your-Writes Gap

Projections update asynchronously, so right after a command succeeds the read model may not reflect it yet. This read-your-writes gap surprises users: they place an order and the list still looks empty for a few milliseconds.

Strategies to manage it:

  • Return the new state from the command so the UI can render optimistically without re-querying.
  • Wait-for-projection: the command returns the event position; the client polls the read model until its checkpoint reaches that position.
  • Design the UX to tolerate brief staleness (spinners, "processing" states).

Never try to make projections synchronous "just to be safe" — you lose the scalability and decoupling that make CQRS worthwhile.

Multiple Projections From One Stream

The real power of CQRS: a single event stream feeds many independent read models, each optimized for a different query. The same OrderPlaced event might update:

  • an OrderList projection for the customer's order history,
  • a DailyRevenue projection for analytics,
  • a SearchIndex projection feeding Elasticsearch.

Each runs with its own handlers and its own checkpoint, so they can be added, rebuilt, or scaled independently.

// One dispatcher fans an event out to many projections
const projections = [];

function register(name, handlers) {
  projections.push({ name, handlers, state: {} });
}

function dispatch(event) {
  for (const p of projections) {
    const h = p.handlers[event.type];
    if (h) h(p.state, event);
  }
}

register('orderCount', {
  OrderPlaced: (s) => { s.count = (s.count ?? 0) + 1; },
});
register('revenue', {
  OrderPlaced: (s) => { s.sum = (s.sum ?? 0) + 1; },
  OrderPaid:   (s, e) => { s.paid = (s.paid ?? 0) + e.total; },
});

dispatch({ type: 'OrderPlaced', orderId: 'o1', total: 40 });
dispatch({ type: 'OrderPaid', orderId: 'o1', total: 40 });
console.log(projections.map((p) => [p.name, p.state]));

Rebuilding a Projection

Because read models are derived data, you can throw them away and recompute them. You rebuild a projection when you:

  • change its schema (add a new column or denormalized field),
  • fix a bug in a handler,
  • add a brand-new read model that needs historical data.

The rebuild recipe:

  • Reset the read store (truncate the table) and reset the checkpoint to 0.
  • Replay the entire stream from the beginning through the handlers.
  • For zero-downtime, build into a new table and atomically swap (blue-green) once it catches up to live.
// Rebuild by replaying the whole log into a fresh projection
function rebuild(eventLog, projection) {
  projection.state = {};          // reset read model
  projection.checkpoint = 0;      // reset position
  for (const event of eventLog) {
    projection.when(event);
    projection.checkpoint = event.position;
  }
  return projection;
}

const log = [
  { type: 'OrderPlaced', orderId: 'o1', total: 10, position: 1 },
  { type: 'OrderPlaced', orderId: 'o2', total: 25, position: 2 },
];
const proj = {
  state: {}, checkpoint: 0,
  when(e) { if (e.type === 'OrderPlaced') this.state[e.orderId] = e.total; },
};
rebuild(log, proj);
console.log(proj.state, 'at', proj.checkpoint);
// { o1: 10, o2: 25 } at 2

Ordering, Failures, and Poison Events

A projector consumes events sequentially to preserve order — a shipped event must never be applied before its placed event. That constraint shapes how you handle errors:

  • On a transient failure (DB blip), retry with backoff; do not advance the checkpoint, so the event is re-applied.
  • On a poison event (one that always throws), you must not block the whole projection forever. Move it to a dead-letter log and alert, so the rest of the stream keeps flowing.
  • Parallelize across partitions (e.g. by aggregate id) when you need throughput, keeping order within each partition.

Keep handlers small and deterministic so failures are rare and reproducible.

Quick Check: The Checkpoint Decision

A projector updates a Postgres read model and stores its checkpoint in a separate table. The service can crash at any moment. Which approach best guarantees correctness?

Recap: Read Models and Projections

You now know how the query side of an event-sourced system is built:

  • A projection is a left fold over the event stream into a denormalized, query-optimized read model.
  • A projector owns handlers, persisted state, and a checkpoint; it subscribes to the store and applies events in order.
  • Save the checkpoint after processing and make handlers idempotent to survive at-least-once redelivery (upserts, per-row position guards).
  • Read models are eventually consistent — manage the read-your-writes gap with optimistic responses or wait-for-projection, never by forcing synchronous projections.
  • One stream can feed many read models, and any of them can be rebuilt by replaying the stream (build-and-swap for zero downtime).
  • Process sequentially per partition; retry transient failures without advancing the checkpoint, and dead-letter poison events.

常见问题解答

「构建读取模型与投影」课时是免费的吗？

是的 — 「构建读取模型与投影」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Node.js Backend Development Bootcamp 课程的其余内容，请升级到 CoddyKit PRO。 Node.js Backend Development Bootcamp 课程共包含 4 节课。

「构建读取模型与投影」这节课中我会学到什么？

从事件流派生经过优化的查询端投影，并使其最终保持一致。 你通过在浏览器中直接运行的动手代码来练习 Node.js Backend Development Bootcamp，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 Node.js Backend Development Bootcamp 需要有经验吗？

无需任何先前经验。CoddyKit 上的 Node.js Backend Development Bootcamp 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。 这是第 3 节课，共 4 节。

「构建读取模型与投影」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 Node.js Backend Development Bootcamp 课中编写并运行代码吗？

能。每节 Node.js Backend Development Bootcamp 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

  1. 事件作为事实来源与仅追加日志
  2. 聚合、命令与领域事件建模
  3. 构建读取模型与投影
  4. 快照、版本控制与事件模式演进
← 返回 Node.js Backend Development Bootcamp