Offline-First Application Patterns

Quick Guide: Reads and writes go to a local database first and the network catches up afterwards. IndexedDB is the store — reached through a wrapper such as Dexie (reactive queries, larger) or idb (thin, ~1.2KB) — and every syncable record carries _syncStatus, _lastModified and _localVersion so the queue knows what is outstanding. Deletes are tombstones, never removals, or a delayed sync resurrects them. The Background Sync API is Chromium-only, so an online listener is the mechanism and background sync the optimisation.

Detailed Resources:

• examples/core.md — syncable entity, repository, sync queue with backoff, connectivity detection, optimistic updates, hooks for status and mutations
• examples/indexeddb.md — Dexie schema and migrations, the idb alternative, multi-tab coordination, quota management
• examples/sync.md — last-write-wins, field-level merge, conflict UI, version vectors, delta cursors, background sync, pull-push, status indicators
• reference.md — storage and conflict-strategy selection, troubleshooting by symptom, performance and security notes

Which path applies

• Read-mostly, writes require connectivity — a local copy for reading, refreshed when online,

with no queue and no conflict resolution. Patterns 1, 4 and 6 are the whole of it.

• Full CRUD offline — mutations queue, conflicts happen, and every pattern here applies. Start

at examples/core.md and pick a resolution strategy from examples/sync.md.

<critical_requirements>

Before writing offline-first code

Treat the local database as authoritative. Every read comes from it and every write lands there before anything is sent, which is what makes the UI answer instantly whatever the connection is doing.

Give every syncable record `_syncStatus`, `_lastModified` and `_localVersion`. Without them there is no way to ask what is outstanding, and no way to tell a conflict from a fresh write.

Delete by writing a `_deletedAt` tombstone. A removed row has nothing left to sync, so the next pull brings the record back.

Queue every mutation and drain the queue on reconnect, with exponential backoff and jitter. A transient 502 is otherwise a permanently lost write, and synchronised retries from many clients are what turn a brief outage into a long one.

Keep an IndexedDB transaction free of any other `await`. The transaction closes as soon as control returns to the event loop with no request pending, so an awaited fetch mid-transaction fails with TRANSACTION_INACTIVE_ERR. Fetch first, then open the transaction.

</critical_requirements>

Auto-detection: IndexedDB, indexedDB.open, IDBDatabase, IDBObjectStore, openDB, DBSchema, Dexie, useLiveQuery, dexie-react-hooks, idb-keyval, sync queue, tombstone, \_syncStatus, \_lastModified, last-write-wins, version vector, offline-first, local-first, navigator.onLine, navigator.storage.persist, BroadcastChannel, QuotaExceededError

Applies to:

• Choosing and shaping a local store that survives a reload and a disconnection
• Modelling sync metadata on records the user edits offline
• Queueing mutations, retrying them, and reporting what is outstanding
• Detecting real connectivity rather than a network interface
• Resolving concurrent edits, from last-write-wins to version vectors
• Showing the user what has saved locally and what has reached the server

Handled elsewhere:

• Intercepting network requests and versioning an HTTP response cache — a response cache stores

what the server said and expires; this skill owns records the user authored

• Precaching an application shell so the app boots without a network
• Collaborative character-level editing, which wants a convergent replicated data type rather than

a queue and a merge

<philosophy>

Philosophy

The network is an enhancement. Local storage is the database, and the server is a peer it reconciles with — which inverts the usual arrangement, where local storage is a cache of the truth.

Two things follow. Writes never block on a request, so the UI responds at disk speed rather than at network speed. And every write becomes a claim that may be contested, which is why sync metadata is foundational rather than an add-on: a record with no version is a record no merge can reason about.

User action
    │
Local database  ←── the single source of truth
    │
UI updates immediately
    │
Sync queue (background)
    │
Server, when reachable
    │
Conflict resolution, if the record moved on both sides
    │
Local database updated

The user's remaining job is trust: they need to see that a change is saved, that it is queued, and that it eventually landed. Sync status is a product surface, not a debugging aid.

</philosophy>

<patterns>

Core patterns

Pattern 1: Syncable entity

Every other pattern reads these fields. Business data and sync metadata stay separate, with the metadata prefixed so a merge can skip it wholesale.

interface SyncableEntity {
  id: string;
  _syncStatus: "synced" | "pending" | "conflicted";
  _lastModified: number;
  _serverTimestamp?: number;
  _localVersion: string;
  _serverVersion?: string;
  _deletedAt?: number; // tombstone
}

Full code: examples/core.md

Pattern 2: Repository

One access point for a collection, so no caller has to remember that a write is two operations. Reads filter tombstones; writes stamp metadata, save locally, then enqueue.

interface DataRepository<T extends SyncableEntity> {
  get(id: string): Promise<T | null>; // null for a tombstone
  getAll(): Promise<T[]>;
  save(item: T): Promise<void>; // local write, then enqueue
  delete(id: string): Promise<void>; // tombstone, then enqueue
  getPendingCount(): Promise<number>;
}

Full code: examples/core.md

Pattern 3: Sync queue with backoff

Exponential delay bounded by a ceiling, plus jitter so reconnecting clients do not arrive together.

const INITIAL_BACKOFF_MS = 1000;
const MAX_BACKOFF_MS = 30_000;
const JITTER_FACTOR = 0.5;

function calculateBackoff(attempt: number): number {
  const delay = Math.min(INITIAL_BACKOFF_MS * 2 ** attempt, MAX_BACKOFF_MS);
  const jitter = delay * JITTER_FACTOR * (Math.random() * 2 - 1);
  return Math.floor(delay + jitter);
}

Full code: examples/core.md — retry limits, ordering by timestamp, dead-letter handling

Pattern 4: Connectivity detection

navigator.onLine reports whether a network interface exists, which is true behind a captive portal and on a router with no upstream. Confirm with a request.

async function checkConnectivity(): Promise<boolean> {
  if (!navigator.onLine) return false; // cheap negative, trustworthy
  try {
    const response = await fetch("/api/health", {
      method: "HEAD",
      cache: "no-store",
    });
    return response.ok;
  } catch {
    return false;
  }
}

Full code: examples/core.md — latency sampling and a "slow" state

Pattern 5: Optimistic update with rollback

Capture the previous value before writing and hand back the undo, so the caller's error path is one call rather than a reconstruction.

async function applyOptimistically<T>(id: string, next: T) {
  const previous = (await localDb.get(id)) ?? null;
  await localDb.put(next);

  return async function rollback() {
    if (previous) await localDb.put(previous);
    else await localDb.delete(id);
  };
}

Full code: examples/core.md

Pattern 6: Connection-aware fetching

Return where the data came from alongside the data, so the UI can say "showing saved data" instead of silently presenting something stale as current.

interface FetchResult<T> {
  data: T;
  source: "network" | "cache";
  timestamp: number;
}

// try network with a timeout, cache the success, fall back to cache on any failure
const response = await fetch(url, { signal: AbortSignal.timeout(TIMEOUT_MS) });

Full code: examples/core.md

Pattern 7: Conflict resolution

Three strategies, in increasing cost. Pick by what the field means, not by what is easiest.

• Last-write-wins — most recent timestamp takes the record. Correct for independent values;

wrong wherever two people edited different fields. examples/sync.md Pattern 17.

• Field-level merge — compare each field against the last common state and conflict only where

both sides moved the same one. examples/sync.md Pattern 18.

• Version vectors — a per-client counter that distinguishes a concurrent edit from a sequential

one without trusting any clock. examples/sync.md Pattern 20.

Where neither side can be discarded, surface the conflict and let the user choose: examples/sync.md Pattern 19.

</patterns>

<red_flags>

Red flags

Breaks at runtime:

• await fetch(...) or await new Promise(setTimeout) inside an IndexedDB transaction — the

transaction has already closed and the next operation throws TRANSACTION_INACTIVE_ERR — do the async work first, then open a short transaction

• A hard delete on a record that has not synced — the next pull resurrects it — write _deletedAt

and sweep tombstones once they are confirmed synced

• Writing without checking the quota — QuotaExceededError surfaces as a failed save with no

warning — check navigator.storage.estimate() and evict before writing

• An unbounded queue — a long offline session exhausts storage, and draining it all at once

exhausts memory — cap the queue and batch the drain

• Awaiting the server before updating the UI — the app is now online-first with extra steps — save

locally, then enqueue

Surprising behaviour:

• navigator.onLine === true means an interface exists, not that anything is reachable
• Safari caps script-writable storage at seven days without user interaction, IndexedDB included;

navigator.storage.persist() helps and an installed app helps more

• navigator.storage.estimate() needs a secure context and answers { usage: 0, quota: 0 }

otherwise

• Background Sync is Chromium-only, so an online listener has to carry the load everywhere else
• IndexedDB has no boolean index type — a compound query matches [userId+completed] against

[userId, 1]

• Dexie's useLiveQuery returns undefined while loading, not null, so a === undefined check

is the loading state

• Two tabs write to one database with no coordination; BroadcastChannel is how they agree

(examples/indexeddb.md Pattern 15)

• Comparing timestamps across devices detects difference, not causality — clock drift makes a

sequential edit look concurrent and the reverse

</red_flags>