RFC: Offline-First Transactions

[KyleAMathews]

(Note: This RFC covers offline mutation persistence and retry - ensuring writes don't get lost when offline. This is distinct from offline data persistence, which involves caching/syncing read data for offline access.)

Summary

TanStack DB applications will persist all mutations to a durable outbox before dispatch, enabling automatic replay when connectivity is restored. The system provides per-key scheduling (parallel across distinct keys, sequential per key), exponential backoff with jitter, failure discrimination via NonRetriableError, and developer hooks for filtering and squashing operations. Optimistic state is restored on restart by replaying persisted transactions, ensuring users never lose work during offline periods.

Background

TanStack DB provides reactive client store with collections, live queries, and optimistic mutations. Currently, when a transaction's mutation function fails, the optimistic state is rolled back and the operation is lost. Users must manually retry operations when connectivity returns.

The framework lacks built-in mechanisms for persisting failed transactions across application restarts, automatically retrying operations when connectivity is restored, or distinguishing between temporary failures (network issues) and permanent failures (validation errors).

Demand for offline-first capabilities spans field service applications, productivity tools, mobile applications, and local-first collaborative systems. Without first-class offline support, developers must either accept data loss during network failures or build complex custom persistence and retry logic outside of TanStack DB.

Problem

Developers using TanStack DB cannot build reliable offline-first applications without significant custom code. This creates three critical problems:

Data Loss During Network Failures: When a transaction's mutation function fails due to network issues, the optimistic updates are rolled back and the user's changes are lost. Users must remember and manually re-enter their data when connectivity returns, leading to frustration and potential data inconsistencies.

No Persistence Across Application Restarts: If the application closes while offline (browser tab closed, mobile app backgrounded, device restarted), any pending operations are permanently lost. There is no mechanism to queue and retry these operations when the application restarts with connectivity.

Inability to Distinguish Failure Types: All errors are treated identically - whether it's a temporary network failure that should be retried or a permanent validation error that will never succeed. This leads to either wasted resources retrying operations that will always fail or premature abandonment of operations that would succeed with retry.

These problems make TanStack DB unsuitable for applications requiring reliable offline operation, forcing developers to either accept data loss or build complex workarounds outside the framework.

Proposal

Core Architecture

Implement an outbox-first persistence system where every offline transaction is stored to durable storage before dispatch. This builds on TanStack DB's existing transaction model by adding persistence and replay capabilities to the current Transaction class.

Outbox Schema

Each outbox transaction extends the existing Transaction format:

{
  id: string,                    // existing transaction ID
  mutatorName: string,           // registry key for offline replay
  mutations: PendingMutation[],  // existing transaction.mutations
  keys: string[],                // derived from all mutation globalKeys
  idempotencyKey: string,        // stable UUID for the entire transaction
  createdAt: Date,               // existing transaction.createdAt
  retryCount: number,            // number of retry attempts
  nextAttemptAt: number,         // next scheduled retry time
  lastError?: SerializedError,   // most recent error details
  metadata?: Record<string, any>, // existing transaction.metadata
  version: 1                     // schema version for future migrations
}

Storage Adapter

The storage layer accepts JavaScript objects and handles serialization internally. Default implementation uses IndexedDB for browsers with fallback to localStorage. Storage quota exceeded errors are thrown to the application for handling.

Intelligent Execution Scheduling

The executor implements per-key scheduling based on PendingMutation.globalKey:

• Parallel execution: Transactions with non-overlapping keys execute concurrently up to maxConcurrency (default 4)
• Sequential execution: Transactions with overlapping keys execute in creation order, blocked on failure
• Key extraction: Uses existing globalKey from mutations (format: ${collection.id}:${itemKey})
• Fairness: When hitting concurrency limits, oldest transactions execute first regardless of key

Retry Policy

Implements infinite retry with exponential backoff:

• Backoff schedule: 1s → 2s → 4s → 8s → 16s → 32s → 60s (capped)
• Jitter: Randomization prevents thundering herd
• Retry-After: Honors server-provided backoff hints
• Fresh timing: Each retry batch recalculates timing; filtered transactions are permanently deleted

Failure Discrimination

• Default behavior: All errors trigger retry with exponential backoff
• NonRetriableError: Transactions are immediately removed from storage and not retried
• Future extensibility: Error classification system can expand for additional error types

Online Detection and Triggers

Retry execution triggers on:

• Application initialization
• navigator.onLine events
• Visibility API changes (tab focus/unfocus)
• Manual developer trigger via notifyOnline()
• Any successful transaction (signals connectivity restored)

Developer Control Hooks

beforeRetry Hook: Called before each retry batch with beforeRetry(transactions[]). Can filter, transform, or squash transactions. Filtered transactions are permanently deleted from storage.

Manual Management: removeFromOutbox(id) for programmatic transaction removal. Optional peekOutbox() for diagnostics.

Optimistic State Restoration

On application restart, persisted transactions replay through the existing transaction system by calling the normal collection operations (insert/update/delete) to restore optimistic UI state. This leverages the existing PendingMutation data structure.

Multi-Tab Coordination

Only one executor runs per origin using:

• Primary: Web Locks API for modern browsers
• Fallback: BroadcastChannel leader election
• Failover: Bounded recovery time when leader tab becomes unresponsive

Leadership transitions handle edge cases gracefully, with delays acceptable for v1 implementation.

API Design

Executor Initialization with Mutator and Collection Registry:

const offline = startOfflineExecutor({
  collections: {
    todos: todoCollection,
    projects: projectCollection,
    // Register all collections that will be used in offline transactions
  },
  mutators: {
    syncTodos: async ({ transaction, idempotencyKey }) => {
      // Handle all mutations in the transaction together
      await api.saveBatch(transaction.mutations, { idempotencyKey })
    },
    updateProject: async ({ transaction, idempotencyKey }) => {
      // Handle project-related mutations 
      await api.updateProject(transaction.mutations, { idempotencyKey })
    },
  },
  maxConcurrency: 4,
  jitter: true,
  beforeRetry: (transactions) => transactions.filter(tx => tx.createdAt > Date.now() - DAY),
  storage: indexedDbAdapter(),
  onUnknownMutator: (name, transaction) => console.warn(`Unknown mutator: ${name}`),
})

Offline Transaction Creation (follows createTransaction pattern):

const tx = offline.createOfflineTransaction({
  mutatorName: 'syncTodos',
  autoCommit: false, // optional, defaults to true
})

tx.mutate(() => {
  // Apply optimistic updates using existing collection APIs
  todoCollection.insert({ id: '123', text: 'Buy milk' })
  todoCollection.update('124', (draft) => { draft.completed = true })
  projectCollection.update('proj1', (draft) => { draft.todoCount += 1 })
})

await tx.commit() // if autoCommit is false

Offline Action Creation (follows createOptimisticAction pattern):

const addTodo = offline.createOfflineAction({
  mutatorName: 'syncTodos',
  onMutate: (text: string) => {
    todoCollection.insert({
      id: crypto.randomUUID(),
      text,
      completed: false
    })
  }
})

// Usage - returns Transaction like createOptimisticAction
const transaction = addTodo('New Todo Item')
await transaction.isPersisted.promise

Automatic Offline Support for Collection Operations

When collections are registered with the offline executor, their existing mutation APIs automatically gain offline capabilities:

Transparent Offline Behavior:

// Existing collection mutation APIs work offline automatically
todoCollection.insert({ id: '123', text: 'Buy milk' })
todoCollection.update('124', draft => draft.completed = true)
todoCollection.delete('125')

Behind the Scenes:

• If todoCollection is registered in the collection registry, these calls automatically create offline transactions
• Uses the collection's registry key ('todos') as the mutatorName
• During replay, the system groups mutations by type and calls the appropriate collection handlers (onInsert/onUpdate/onDelete)

Explicit Transactions (for custom mutators):

// Only needed when using custom mutators that handle multiple collections
const tx = offline.createOfflineTransaction({
  mutatorName: 'syncTodos' // Custom mutator for complex operations
})

Integration with Existing System

This builds directly on the existing transaction system:

• Reuses Transaction class: Extends rather than replaces current transaction model
• Preserves PendingMutation: Uses existing mutation data structure for persistence
• Maintains optimistic updates: Leverages existing optimistic state management
• Zero-config offline: Existing collection operations work offline when collections are registered
• Progressive enhancement: Developers can add explicit mutators for advanced use cases

Definition of success

This proposal succeeds when developers can build reliable offline-first applications with TanStack DB without custom persistence logic. Success metrics include:

Functional Requirements Met:

• Failed transactions persist across application restarts with full mutation data intact
• Automatic retry system handles network failures with exponential backoff and intelligent scheduling
• NonRetriableError prevents infinite retry of permanent failures
• Per-key scheduling enables parallel execution while maintaining consistency
• Multi-tab coordination prevents duplicate retry execution

Developer Experience Goals:

• APIs follow existing TanStack DB patterns (createTransaction, createOptimisticAction)
• Minimal configuration required for basic offline functionality
• Clear upgrade path from existing transaction usage
• Comprehensive error handling and debugging capabilities

Performance Targets:

• Outbox operations add minimal overhead to normal transaction flow
• Parallel retry execution maximizes throughput when connectivity returns
• Storage operations remain responsive under typical offline workloads
• Memory usage scales reasonably with outbox size

Reliability Standards:

• Zero data loss during offline periods and application restarts
• Graceful handling of storage quota and corruption scenarios
• Predictable behavior during network transitions and multi-tab usage
• Clear failure modes with actionable error messages

The feature succeeds when developers can add offline capabilities to existing TanStack DB applications with minimal code changes while maintaining the framework's reactive performance characteristics.

Comparison with Existing Solutions

Dimension	TanStack DB Offline Transactions	TanStack Query (persisted/paused mutations)	Redux-Offline (Outbox)	Replicache
Core model	Outbox-first (persist before dispatch); replay on init/online	Persist paused mutations; resume if default mutationFn available	Outbox; queue of actions flushed when online	Local-first DB; named mutators + args; server sync
Mutation representation	Transaction with PendingMutation[] bound by mutator registry; no closures	Function ref + variables (functions not serializable → needs default fn)	Action object; app reducer handles effects	mutatorName + JSON args; deterministic; re-runnable
Idempotency	Auto-generated idempotencyKey per transaction; optional usage	None built-in; app could implement	Not built-in; app concern	Strongly encouraged; assumption in design
Parallelism / ordering	Parallel across keys, serial per key; derived from globalKey	Per-mutation; no key-aware scheduler	Serial unless you build custom middleware	Mutator stream; server-side ordering by version/lsn
Keying	Auto-derived from existing PendingMutation.globalKey	N/A (no per-key scheduler)	N/A	Per-doc / per-space keys; CRDT-friendly patterns
Retry policy	Infinite, expo backoff + jitter, honors Retry-After	Retry via Query's mechanisms; limited backoff control	Configurable backoff	Client retries; server reconciliation
Failure taxonomy (v1)	Retry by default; NonRetriableError drops transaction	App-defined	App-defined	App-defined conflicts; server wins after push/pull
Optimistic on restart	Yes: replay transactions to restore UI state immediately	Partial via cache rehydrate, but no cross-reload optimistic replay	Usually app-specific; often no	Yes (local DB is source of truth)
Multi-tab leader election	Yes: Web Locks → BroadcastChannel fallback	No (each tab manages its own)	Usually no (you add it)	Yes (LeaderElection via broadcast-channel)
Service Worker / BG Sync	Out of scope v1 (can layer later)	N/A	Optional community patterns	N/A (not required)
Storage	IndexedDB/localStorage adapter; async	Persist Query cache + paused mutations (IndexedDB)	Redux store + storage (often IndexedDB)	IndexedDB (browser) + server
Dev hooks	beforeRetry(transactions[]), removeFromOutbox, optional peekOutbox()	Mutation lifecycle callbacks	Configurable offline/online/commit hooks	Custom mutators; pull/push hooks
Conflict handling	App-defined (mutator layer + beforeRetry rewrite/squash)	App-defined per mutation	App-defined reducers	Built-in patterns (server authoritative; app merges)
API shape	startOfflineExecutor({ mutators }), offline.createOfflineTransaction, offline.createOfflineAction	persistQueryClient + mutation defaults	Higher-order store enhancer + config	rep.mutate.<name>(args); server sync protocol
Philosophy fit	Extend existing TanStack DB transactions with durable outbox semantics	Online-first; offline is a pause/resume convenience	Offline-capable apps with Redux	Full local-first collaboration model
Integration with TanStack DB	Native: extends existing Transaction and PendingMutation	External: would need custom integration layer	External: would need custom integration layer	External: would need custom integration layer

Key Differentiators

vs TanStack Query: TanStack DB Offline Transactions provides key-aware scheduling and true transaction persistence across restarts, rather than just pausing individual mutations. The integration is native since it extends the existing transaction system.

vs Redux-Offline: Built-in parallelism via automatic key derivation eliminates the need for custom middleware. The system integrates directly with TanStack DB's reactive collections and optimistic updates.

vs Replicache: Focused on write-path reliability rather than full local-first database replacement. Developers keep their existing backend architecture while gaining offline write resilience.

Positioning: TanStack DB Offline Transactions extends your existing TanStack DB transaction workflow with durable outbox semantics and intelligent retry scheduling. It doesn't replace your backend, conflict strategy, or sync engine - it makes your existing write path safe under flaky networks and app restarts, with minimal API surface area.

[thruflo]

Looks great 👍

A few observations:

1. multi-tab leader election is a non-trivial concern to casually mix in there
2. could the explicit transactions use the existing createTransaction and createOptimisticAction APIs with an offline: true flag or other config method? Rather than introducing the new function api?
3. where does the user handle write rejection? Ie once the optimistic state is applied locally and the transaction is persisted and being retried, how does the user control matching on sync after the write and handling write failure? Is it just in the mutation handler that’s being retried for you, so everything is exactly the same as the non-offline case? If so, that’s awesome but there is a slight difference because the write can be being rejected a long time after it was made, which may be more surprising.

[KyleAMathews]

1. we can't really get away from this — once data is somewhere that multiple tabs can access, we need to gate access to it
2. not really — you need to tie them the offline executor which tracks mutationFns & collections so we can safely hydrate persisted offline txs — generally speaking though the APIs are identical
3. You throw a NonRetriableError. You match on sync like normal so yeah, everything is the same other than move the mutationFn/Collection bits to named registries as we can't persist those.

the write can be being rejected a long time after it was made, which may be more surprising.

yes, the docs will strongly urge people not to use this unless they really want to think through these things. "PLEASE DO NOT USE THIS API" 😆

[samwillis]

When collections are registered with the offline executor, their existing mutation APIs automatically gain offline capabilities

I like that this is possible, be as @thruflo mentions with the existing api, maybe a flag would be nice to turn on/off. And I assume a collection can therefore only be registered with one offline mutation executor.

Multi-Tab Coordination

This is going to be complex, WebLocks work well(ish) for leader election, but as we have seen with PGlite and its leader election it can be quite flaky. BroadcastChannel is not the way to do this. "Web Locks → BroadcastChannel fallback", I doubt we want the complexity of the fallback. Also note that we will likely want to move the leader to the current focused tab too otherwise the setTimout on throttling will not work as they are themselves throttled when in the background.

Reading this through, my biggest concern is that without also having the sync managed through a leader election, sharing the offline mutations between tabs could results in some strange broken state or race conditions. A transaction that is awaiting a sync back, could be "applied" by the leader tab before the sync has caught up on other tabs.

[KyleAMathews]

sharing the offline mutations between tabs could results in some strange broken state or race conditions. A transaction that is awaiting a sync back, could be "applied" by the leader tab before the sync has caught up on other tabs.

Hmm yeah, both tabs trying to do do offline mutations seems... tricky. I'm kinda thinking we just avoid this problem (for now perhaps) by only allowing one tab to do offline mutations. The second one can do online writes but not offline.

[samwillis]

Should also say other than that its looking good!

[justjake]

I don't understand the description of keys. Can you describe the "global keys" concept in more detail or point me towards something that specifies it?

My main apprehension as currently described is that causal dependencies between transactions may not be reflected in these keys. For example in Notion's transaction system, we can have causal dependencies between transactions:

Transaction 1:
  CREATE { id: X, parentId: null }

Transaction 2:
  CREATE { id: Y, parentId: X } // fails if X does not exist

[KyleAMathews]

Ooo yeah — dependencies between unrelated transactions isn't something we've really worried about to date w/o online writes as a) you'd generally group those in the same transaction or b) if you didn't want these two transactions together, you'd need to await tx 1 to finish persisting and then start tx 2 w/ your own dependency tracking thing. Generally not an issue given persisting is fast when online.

But for offline we need something better. I asked ChatGPT and it suggested that offline transactions could have a provides and requires array of entity ids and that the persister could then create a DAG to drive the ordering of persisting. Which seems workable.

[justjake]

Perhaps "provides" and "requires" works for existence constraint, but data models that have more exacting constraints may exist. The most notable is permission constraints.

[KyleAMathews]

Yeah 🤔 Perhaps beforeRetry could let you return explicit dependencies. Or you could opt-into a global queue where you persist things sequentially by creation time.

[justjake]

I’d require the developer explicitly pick between total global order, order per collection, or the explicit dependency api y’all have with the keys thing, provides, when setting up offline transaction queue. i think application dev should have to consider this because it’s not immediately obvious for someone doing offline for the first time and getting it wrong means you end up shredding the user’s data.

as an aside, i think the provides/requires is very easy to get wrong / forget as the team building the app grows, and again getting it wrong will shred user data if you have any constraints. If i was building a new app from scratch i would definitely avoid using that semantics.

[ignatz]

Great RFC. Thanks for looking into this ❤️

A few questions:

1. "This RFC covers offline mutation persistence and retry - ensuring writes don't get lost when offline. This is distinct from offline data persistence, which involves caching/syncing read data for offline access." - I love that you scope this down. Just for my curiosity, we do expect that "offline first" will also eventually require "offline persistence"? Absolutely no action requested here, just wanting to make sure, I'm not misunderstanding.
2. You talk about how to schedule the various mutations in your backlog during replay. However, I wasn't quite sure how this is coordinated with, e.g. new changes, refetches....? Maybe you could have an example, was what happens when you come back online, e.g. do you start receiving updates/refreshing right away or first replay the backlog? Is there some per-key coordination?
3. Ultimately, I'm most interested in the conflict resolution. When you say: "App-defined (mutator layer + beforeRetry rewrite/squash)", maybe you could talk a bit more through "the life of a conflict", e.g. a mutation gets replayed, it's rejected by the server based of an out of order write, how would a merge/rewrite/squash in the beforeRetry hook look like?

[KyleAMathews]

Oh sure, that's how things already work. Mutations have a changes object and that's what we document — so in your case, the mutationFn would have a changes object of { title: 'new title' } which you'd pass onto your API.

Also if you have a write that's dependent on read e.g. a counter, you'd want to call /api/increment-count or whatever not send { count: 100 }

[KyleAMathews]

Or if you mean that optimistic updates should be merged over fresh server data — we've thought about that but it does seem tricky to manage as an optimistic update might only touch a few fields but the server might touch more so the synced server data wouldn't merge cleanly.

[justjake]

Or if you mean that optimistic updates should be merged over fresh server data

yeah that’s what i mean. In many cases this is what you want, and the experience can feel pretty bad if your transaction queue is slow but you’re online and trying to collaborate; it’s edge case maybe but we see this when user comes back online after a while and they have a deep queue.

[KyleAMathews]

user comes back online after a while and they have a deep queue.

Ah! Yeah with offline-first transactions, exposing some sort of option on collections to merge optimistic updates into synced objects (instead of replacing) makes a lot of sense for improving the responsiveness on startup.

[ignatz]

Ah! Yeah with offline-first transactions, exposing some sort of option on collections to merge optimistic updates into synced objects (instead of replacing) makes a lot of sense for improving the responsiveness on startup.

That sounds great. This is what I was wondering or hoping for when fumbling my words around "coordination between syncing and replaying"

Yeah that's the plan though far fewer web apps IME need offline persistence as almost nobody uses Service Workers which is required to actual load the app code offline. Native apps do more often need offline persistence.

Makes sense (I'm certainly coming from the "native apps" side, e.g. react native, tauri, ... maybe less common but maybe where offline-first is most desirable 🤷‍♀️ )

👍

[gruckion]

TODO, consider how this will work with Partitioned collections #315
.

[yunusyuksel]

That looks great
This is the only feature currently preventing me from switching to TanStack DB for my offline-first application, and I imagine it affects many other developers building offline-first apps as well.

[daveycodez]

Is there any ETA on this feature? I'm building a complex multi-layer system with RxDB and this would totally eliminate the need for that

[KyleAMathews]

It should be ready for merging. Just needs another review.

[yunusyuksel]

@KyleAMathews I see the PR is merged.

I am building react native application. Would it work also with react native?

[KyleAMathews]

not atm. See my comment here — someone will need to pickup this work #559 (comment)