Custom Conflict Resolution

The default behavior is “last write wins per field”. Updates to different fields on the same record don’t conflict with each other. The server processes operations in the order received, so if two users modify the same field, the last update to reach the server wins. For most apps, this works fine. But some scenarios demand more complex conflict resolution strategies.

When You Might Need Custom Conflict Resolution

Retail inventory: Two clerks ring up the same item while offline. You need to subtract both quantities, not replace one count with the other. Healthcare records: A doctor updates diagnosis while a nurse updates vitals on the same patient record. Both changes matter, you can’t lose either. Order workflows: Once an order ships, it should lock. Status must progress logically (pending → processing → shipped), not jump around randomly. Collaborative documents: Multiple people edit different paragraphs simultaneously. Automatic merging prevents losing anyone’s work.

How Data Flows Through PowerSync

Understanding the data flow helps you decide where to implement conflict resolution.

Client to Backend

When a user updates data in your app:

Client writes to local SQLite - Changes happen instantly, even offline
PowerSync queues the operation - Stored in the upload queue
Client sends operation(s) to your backend - Your uploadData function processes it
Backend writes to source database - Postgres, MySQL, MongoDB etc.

Backend to Client

When data changes on the server:

Source database updates - Direct writes or changes from other clients
PowerSync Service detects changes - Through replication stream
Clients download updates - Based on their Sync Streams (or legacy Sync Rules)
Local SQLite updates - Changes merge into the client’s database

Conflicts arise when: Multiple clients modify the same row (or fields) before syncing, or when a client’s changes conflict with server-side rules.

Understanding Operations & `CrudEntry`

PowerSync tracks three operation types:

PUT - Creates new row or replaces entire row (includes all non-null columns)
PATCH - Updates specific fields only (includes ID + changed columns)
DELETE - Removes row (includes only ID)

`CrudEntry` Structure

When your uploadData receives transactions, each one has this structure:

interface CrudEntry {
  clientId: number;                      // Auto-incrementing client ID
  id: string;                            // ID of the changed row
  op: UpdateType;                        // 'PUT' | 'PATCH' | 'DELETE'
  table: string;                         // Table name
  opData?: Record<string, any>;          // Changed column values (optional)
  transactionId?: number;                // Groups ops from same transaction
  metadata?: string;                     // Custom metadata (trackMetadata)
  trackPrevious?: Record<string, any>;  // Previous values (trackPrevious)
}

What Your Backend Receives

Client-side connector sends:

// uploadData in your client connector
async uploadData(database) {
  const transaction = await database.getNextCrudTransaction();
  if (!transaction) return;
  
  // Send to your backend API
  await fetch('https://yourapi.com/data', {
    method: 'POST',
    body: JSON.stringify({
      batch: transaction.crud  // Array of CrudEntry objects
    })
  });
  
  await transaction.complete();
}

The following structure is only received by the backend if the transactions are not mutated in your client’s uploadData function

Backend API receives:

{
  "batch": [
    {
      "op": "PATCH",
      "table": "todos",
      "id": "44f21466-d031-11f0-94bd-62f5a66ac26c",
      "opData": {
        "completed": 1,
        "completed_at": "2025-12-03T10:20:04.658Z",
        "completed_by": "c7b8cc68-41dd-4643-b559-66664ab6c7c5"
      }
    }
  ]
}

Operations are idempotent - your backend may receive the same operation multiple times. Use clientId and the operation’s ID to detect and skip duplicates.

Implementation Examples

The following examples demonstrate the core logic and patterns for implementing conflict resolution strategies. All client-side code is written for React/Web applications, backend examples use Node.js, and database queries target Postgres. While these examples should work as-is, they’re intended as reference implementations, focus on understanding the underlying patterns and adapt them to your specific stack and requirements.

Strategy 1: Timestamp-Based Detection

The idea is simple: add a modified_at timestamp to each row. When a client updates a row, compare their timestamp to the one in the database. If theirs is older, someone else changed the row while they were offline, so you treat it as a conflict. This is great for quick staleness checks. You are not merging changes, just stopping outdated writes, similar to noticing a Google Doc changed while you were editing a local copy. The only real catch is clock drift. If server and client clocks are out of sync, you can get false conflicts. And if clients generate timestamps themselves, make sure they all use the same timezone.

Database Schema

Source database (Postgres):

CREATE TABLE tasks (
  id UUID PRIMARY KEY,
  title TEXT,
  status TEXT,
  modified_at TIMESTAMPTZ NOT NULL DEFAULT NOW()
);

-- Auto-update timestamp on every change
CREATE OR REPLACE FUNCTION update_modified_at()
RETURNS TRIGGER AS $$
BEGIN
  NEW.modified_at = NOW();
  RETURN NEW;
END;
$$ LANGUAGE plpgsql;

CREATE TRIGGER tasks_modified_at
  BEFORE UPDATE ON tasks
  FOR EACH ROW
  EXECUTE FUNCTION update_modified_at();

Backend Conflict Detection

Backend API (Node.js):

async function handleUpdate(operation, userId) {
  const { id, opData } = operation;
  const clientModifiedAt = opData.modified_at;
  
  // Get current server state
  const result = await db.query(
    'SELECT * FROM tasks WHERE id = $1',
    [id]
  );
  
  if (!result.rows[0]) {
    // Row was deleted by another client
    console.log(`Conflict: Row ${id} deleted`);
    return { conflict: 'row_deleted' };
  }
  
  const serverModifiedAt = result.rows[0].modified_at;
  
  // Client's version is older than server's
  if (new Date(clientModifiedAt) < new Date(serverModifiedAt)) {
    console.log(`Conflict: Stale update for ${id}`);
    return {
      conflict: 'stale_update',
      serverVersion: result.rows[0],
      clientVersion: opData
    };
  }
  
  // No conflict - apply update
  await db.query(
    'UPDATE tasks SET title = $1, status = $2 WHERE id = $3',
    [opData.title, opData.status, id]
  );
  
  return { success: true };
}

Timestamps can be unreliable if servers have clock skew. Additionally, if clients are writing timestamps (rather than letting the database generate them), ensure all clients use the same timezone/localization as the server. For critical data, use sequence numbers instead.

Strategy 2: Sequence Number Versioning

Instead of timestamps, you can use a version number that increments on every change. It works like a counter on the row. Each time someone updates it, the version increases by one. When a client sends an update, they include the version they last saw. If it doesn’t match the current version in the database, another update happened and you reject the write. This avoids clock drift entirely because the database manages the counter, so clients can’t get out of sync. The tradeoff is that it’s all or nothing. You can’t merge simultaneous edits to different fields. You only know that the row changed, so the update is rejected. Use this when you want strong conflict detection and are fine asking users to refresh and redo their edits rather than risking corrupted data.

Database Schema

Source database (Postgres):

CREATE TABLE documents (
  id UUID PRIMARY KEY,
  content TEXT,
  version BIGSERIAL NOT NULL
);

Backend Conflict Detection

Backend API (Node.js):

async function handleUpdateWithVersion(operation) {
  const { id, opData } = operation;
  const clientVersion = opData.version;
  
  const result = await db.query(
    'SELECT version FROM documents WHERE id = $1',
    [id]
  );
  
  if (!result.rows[0]) {
    return { conflict: 'row_deleted' };
  }
  
  const serverVersion = result.rows[0].version;
  
  // Client's version doesn't match server
  if (clientVersion !== serverVersion) {
    return {
      conflict: 'version_mismatch',
      expected: serverVersion,
      received: clientVersion
    };
  }
  
  // Update and increment version atomically
  await db.query(
    'UPDATE documents SET content = $1, version = version + 1 WHERE id = $2',
    [opData.content, id]
  );
  
  return { success: true };
}

Strategy 3: Field-Level Last Write Wins

Here things get more fine-grained. Instead of tracking changes for the whole row, you track them per field. If one user updates the title and another updates the status, both changes can succeed because they touched different fields. You store a timestamp for each field you care about. When an update comes in, you compare the client’s timestamp for each field to what’s in the database and only apply the fields that are newer. This allows concurrent edits to coexist as long as they are not modifying the same field. The downside is extra complexity. You end up with more timestamp columns, and your backend has to compare fields one by one. But for apps like task managers or form builders, where different parts of a record are often edited independently, this avoids a lot of unnecessary conflicts.

Database Schema

Source database (Postgres):

CREATE TABLE tasks (
  id UUID PRIMARY KEY,
  title TEXT,
  title_modified_at TIMESTAMPTZ,
  status TEXT,
  status_modified_at TIMESTAMPTZ,
  assignee TEXT,
  assignee_modified_at TIMESTAMPTZ
);

Client Schema with Metadata

Client schema:

const tasks = new Table(
  {
    title: column.text,
    status: column.text,
    assignee: column.text,
    // Store per-field timestamps in metadata
  },
  {
    trackMetadata: true  // Enables _metadata column
  }
);

Client Updates with Timestamps

Client code:

await powerSync.execute(
  'UPDATE tasks SET title = ?, _metadata = ? WHERE id = ?',
  [
    'Updated title',
    JSON.stringify({
      title_modified_at: new Date().toISOString(),
      status_modified_at: existingTask.status_modified_at // Keep existing
    }),
    taskId
  ]
);

Backend Field-Level Resolution

Backend API (Node.js):

async function fieldLevelLWW(operation) {
  const { id, opData, metadata } = operation;
  const timestamps = metadata ? JSON.parse(metadata) : {};
  
  // Get current field timestamps from database
  const result = await db.query(
    'SELECT title_modified_at, status_modified_at, assignee_modified_at FROM tasks WHERE id = $1',
    [id]
  );
  
  if (!result.rows[0]) {
    return { conflict: 'row_deleted' };
  }
  
  const currentTimestamps = result.rows[0];
  const updates = [];
  const values = [];
  let paramCount = 1;
  
  // Check each field that was updated
  for (const [field, value] of Object.entries(opData)) {
    if (field === 'id') continue;
    
    const clientTimestamp = timestamps[`${field}_modified_at`];
    const serverTimestamp = currentTimestamps[`${field}_modified_at`];
    
    // Only update if client's version is newer (or server has no timestamp)
    if (!serverTimestamp || 
        (clientTimestamp && new Date(clientTimestamp) > new Date(serverTimestamp))) {
      updates.push(`${field} = $${paramCount}`);
      updates.push(`${field}_modified_at = $${paramCount + 1}`);
      values.push(value, clientTimestamp);
      paramCount += 2;
    }
  }
  
  if (updates.length > 0) {
    values.push(id);
    await db.query(
      `UPDATE tasks SET ${updates.join(', ')} WHERE id = $${paramCount}`,
      values
    );
  }
  
  return { success: true };
}

Strategy 4: Business Rule Validation

Sometimes conflicts aren’t about timing at all, they’re about your business rules. Maybe an order that has shipped can’t be edited, or a status can’t jump from pending to completed without hitting processing or prices can only change with manager approval. This approach isn’t about catching concurrent edits. It’s about enforcing valid state transitions. You look at the current state in the database, compare it to what the client wants, and decide whether that move is allowed. This is where your domain rules live. The logic becomes the gatekeeper that blocks changes that don’t make sense. You can also layer it with other methods: check timestamps first, then validate your business rules, and only then apply the update.

Backend with Business Rules

Backend API (Node.js):

async function validateOrderUpdate(operation) {
  const { id, opData } = operation;
  
  const result = await db.query(
    'SELECT * FROM orders WHERE id = $1',
    [id]
  );
  
  if (!result.rows[0]) {
    return { conflict: 'row_deleted' };
  }
  
  const serverOrder = result.rows[0];
  
  // Rule 1: Shipped orders are immutable
  if (serverOrder.status === 'shipped' || serverOrder.status === 'completed') {
    return {
      conflict: 'order_locked',
      message: 'Cannot modify shipped or completed orders'
    };
  }
  
  // Rule 2: Validate status transitions
  const validTransitions = {
    'pending': ['processing', 'cancelled'],
    'processing': ['shipped', 'cancelled'],
    'shipped': ['completed'],
    'completed': [],
    'cancelled': []
  };
  
  if (opData.status && 
      !validTransitions[serverOrder.status]?.includes(opData.status)) {
    return {
      conflict: 'invalid_transition',
      message: `Cannot change status from ${serverOrder.status} to ${opData.status}`
    };
  }
  
  // Rule 3: Price changes need approval flag
  if (opData.price !== undefined && 
      opData.price !== serverOrder.price && 
      !opData.manager_approved) {
    return {
      conflict: 'approval_required',
      message: 'Price changes require manager approval'
    };
  }
  
  // Rule 4: Stock level must be positive
  if (opData.quantity !== undefined && opData.quantity < 0) {
    return {
      conflict: 'invalid_quantity',
      message: 'Quantity cannot be negative'
    };
  }
  
  // All validations passed
  const updateFields = [];
  const updateValues = [];
  let paramCount = 1;
  
  for (const [field, value] of Object.entries(opData)) {
    if (field === 'id') continue;
    updateFields.push(`${field} = $${paramCount}`);
    updateValues.push(value);
    paramCount++;
  }
  
  updateValues.push(id);
  await db.query(
    `UPDATE orders SET ${updateFields.join(', ')} WHERE id = $${paramCount}`,
    updateValues
  );
  
  return { success: true };
}

Strategy 5: Server-Side Conflict Recording

Sometimes you can’t automatically fix a conflict. Both versions might be valid, and you need a human to choose. In those cases you record the conflict instead of picking a winner. You save both versions in a write_conflicts table and sync that back to the client so the user can decide. The flow is simple: detect the conflict, store the client and server versions, surface it in the UI, and let the user choose or merge. After they resolve it, you mark the conflict as handled. This is the safest option for high-stakes data where losing either version isn’t acceptable, like medical records, legal documents, or financial entries. The tradeoff is extra UI work and shifting the final decision to the user.

Step 1: Create Conflicts Table

Source database (Postgres):

CREATE TABLE write_conflicts (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  table_name TEXT NOT NULL,
  row_id UUID NOT NULL,
  conflict_type TEXT NOT NULL,
  client_data JSONB NOT NULL,
  server_data JSONB NOT NULL,
  resolved BOOLEAN DEFAULT FALSE,
  user_id UUID NOT NULL,
  created_at TIMESTAMPTZ DEFAULT NOW()
);

Step 2: Sync Conflicts to Clients

Sync Streams / Sync Rules:

Sync Streams
Sync Rules (Legacy)

config:
  edition: 3
streams:
  user_data:
    queries:
      - SELECT * FROM tasks WHERE user_id = auth.user_id()
      - SELECT * FROM write_conflicts WHERE user_id = auth.user_id() AND NOT resolved

bucket_definitions:
  user_data:
    parameters:
      - SELECT request.user_id() as user_id
    data:
      - SELECT * FROM tasks WHERE user_id = bucket.user_id
      - SELECT * FROM write_conflicts WHERE user_id = bucket.user_id AND resolved = FALSE

Step 3: Record Conflicts in Backend

Backend API (Node.js):

async function handleUpdateWithConflictRecording(operation, userId) {
  const { id, opData } = operation;
  
  const result = await db.query(
    'SELECT * FROM tasks WHERE id = $1',
    [id]
  );
  
  if (!result.rows[0]) {
    return { conflict: 'row_deleted' };
  }
  
  const serverData = result.rows[0];
  const clientModifiedAt = opData.modified_at;
  const serverModifiedAt = serverData.modified_at;
  
  // Detect conflict
  if (new Date(clientModifiedAt) < new Date(serverModifiedAt)) {
    // Record for manual resolution
    await db.query(
      `INSERT INTO write_conflicts 
       (table_name, row_id, conflict_type, client_data, server_data, user_id)
       VALUES ($1, $2, $3, $4, $5, $6)`,
      [
        'tasks',
        id,
        'update_conflict',
        JSON.stringify(opData),
        JSON.stringify(serverData),
        userId
      ]
    );
    
    // Don't apply the update - let user resolve it
    return { conflict: 'recorded' };
  }
  
  // No conflict - apply update
  await db.query(
    'UPDATE tasks SET title = $1, status = $2 WHERE id = $3',
    [opData.title, opData.status, id]
  );
  
  return { success: true };
}

Step 4: Build Resolution UI

Client UI (React):

import { useQuery } from '@powersync/react';
import { powerSync } from './db'

function ConflictResolver() {
  const { data: conflicts } = useQuery(
    'SELECT * FROM write_conflicts WHERE resolved = FALSE'
  );
  
  const resolveConflict = async (
    conflictId: string, 
    useClientVersion: boolean
  ) => {
    const conflict = conflicts.find(c => c.id === conflictId);
    const clientData = JSON.parse(conflict.client_data);
    
    if (useClientVersion) {
      // Reapply client's changes
      const fields = Object.keys(clientData).filter(k => k !== 'id');
      const placeholders = fields.map(() => '?').join(', ');
      const updates = fields.map(f => `${f} = ?`).join(', ');
      
      await powerSync.execute(
        `UPDATE ${conflict.table_name} SET ${updates} WHERE id = ?`,
        [...fields.map(f => clientData[f]), conflict.row_id]
      );
    }
    // If using server version, it's already applied
    
    // Mark as resolved
    await powerSync.execute(
      'UPDATE write_conflicts SET resolved = TRUE WHERE id = ?',
      [conflictId]
    );
  };
  
  if (!conflicts || conflicts.length === 0) {
    return null;
  }
  
  return (
    <div className="conflict-banner">
      <h3>⚠️ {conflicts.length} Conflict(s) Need Your Attention</h3>
      {conflicts.map(conflict => {
        const clientData = JSON.parse(conflict.client_data);
        const serverData = JSON.parse(conflict.server_data);
        
        return (
          <div key={conflict.id} className="conflict-card">
            <p>
              Conflict in <strong>{conflict.table_name}</strong> 
              from {new Date(conflict.created_at).toLocaleString()}
            </p>
            
            <div className="conflict-options">
              <div className="version-option">
                <h4>Your Changes:</h4>
                <ul>
                  {Object.entries(clientData).map(([key, value]) => (
                    <li key={key}>
                      <strong>{key}:</strong> {JSON.stringify(value)}
                    </li>
                  ))}
                </ul>
                <button onClick={() => resolveConflict(conflict.id, true)}>
                  Keep My Version
                </button>
              </div>
              
              <div className="version-option">
                <h4>Server Version:</h4>
                <ul>
                  {Object.entries(serverData).map(([key, value]) => (
                    <li key={key}>
                      <strong>{key}:</strong> {JSON.stringify(value)}
                    </li>
                  ))}
                </ul>
                <button onClick={() => resolveConflict(conflict.id, false)}>
                  Keep Server Version
                </button>
              </div>
            </div>
          </div>
        );
      })}
    </div>
  );
}

Strategy 6: Change-Level Status Tracking

This approach works differently. Instead of merging everything in one atomic update, you log each field change as its own row in a separate table. If a user edits the title of a task, you still apply an optimistic update to the main table, but you also write a row to a field_changes table that records who changed what and to which value. Your backend then processes these changes asynchronously. Each one gets a status like pending, applied, or failed. If a change fails validation, you mark it as failed and surface the error in the UI. The user can see exactly which fields succeeded and which didn’t, and retry the failed ones without resubmitting everything. This gives you excellent visibility. You get a clear history of every change, who made it, and when it happened. The cost is extra writes, since every field update creates an additional log entry. But for compliance-heavy systems or any app that needs detailed auditing, the tradeoff could be worth it. The implementation below shows the full version with complete status tracking. If you don’t need all that complexity, see the simpler variations at the end of this section.

Step 1: Create Change Log Table

Source database (Postgres):

CREATE TABLE field_changes (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  table_name TEXT NOT NULL,
  row_id UUID NOT NULL,
  field_name TEXT NOT NULL,
  new_value TEXT,
  status TEXT DEFAULT 'pending', -- 'pending', 'applied', 'failed'
  error_message TEXT,
  user_id UUID NOT NULL,
  created_at TIMESTAMPTZ DEFAULT NOW()
);

Step 2: Client Writes to Both Tables

Client code:

async function updateTaskField(
  taskId: string,
  field: string,
  newValue: any,
  powerSync: PowerSyncDatabase
) {
  
  await powerSync.writeTransaction(async (tx) => {
    // Optimistic update to main table
    await tx.execute(
      `UPDATE tasks SET ${field} = ? WHERE id = ?`,
      [newValue, taskId]
    );
    
    // Log the change for server tracking
    await tx.execute(
      `INSERT INTO field_changes 
       (table_name, row_id, field_name, new_value, user_id)
       VALUES (?, ?, ?, ?, ?)`,
      ['tasks', taskId, field, String(newValue), getCurrentUserId()]
    );
  });
}

Step 3: Backend Processes Changes

Backend API (Node.js):

async function processFieldChanges() {
  const result = await db.query(
    `SELECT * FROM field_changes 
     WHERE status = 'pending' 
     ORDER BY created_at ASC
     LIMIT 100`
  );
  
  for (const change of result.rows) {
    try {
      // Validate the change
      const isValid = await validateFieldChange(change);
      
      if (!isValid.valid) {
        await db.query(
          `UPDATE field_changes 
           SET status = 'failed', error_message = $1 
           WHERE id = $2`,
          [isValid.reason, change.id]
        );
        continue;
      }
      
      // Apply to main table
      await db.query(
        `UPDATE ${change.table_name} 
         SET ${change.field_name} = $1 
         WHERE id = $2`,
        [change.new_value, change.row_id]
      );
      
      // Mark as applied
      await db.query(
        `UPDATE field_changes SET status = 'applied' WHERE id = $1`,
        [change.id]
      );
      
    } catch (error) {
      await db.query(
        `UPDATE field_changes 
         SET status = 'failed', error_message = $1 
         WHERE id = $2`,
        [error.message, change.id]
      );
    }
  }
}

async function validateFieldChange(change) {
  // Example validation
  if (change.field_name === 'price' && parseFloat(change.new_value) < 0) {
    return { valid: false, reason: 'Price cannot be negative' };
  }
  
  return { valid: true };
}

Step 4: Display Change Status

Client UI (React):

function TaskEditor({ taskId }: { taskId: string }) {
  const { data: pendingChanges } = useQuery(
    `SELECT * FROM field_changes 
     WHERE row_id = ? 
     AND table_name = 'tasks'
     AND status IN ('pending', 'failed')
     ORDER BY created_at DESC`,
    [taskId]
  );
  
  const retryChange = async (changeId: string) => {
    await powerSync.execute(
      'UPDATE field_changes SET status = ? WHERE id = ?',
      ['pending', changeId]
    );
  };
  
  return (
    <div>
      {/* Your task editing form */}
      
      {pendingChanges && pendingChanges.length > 0 && (
        <div className="sync-status">
          {pendingChanges.map(change => (
            <div key={change.id} className={`status-${change.status}`}>
              {change.status === 'pending' && (
                <span>⏳ Syncing {change.field_name}...</span>
              )}
              {change.status === 'failed' && (
                <div>
                  <span>
                    ❌ Failed to update {change.field_name}: {change.error_message}
                  </span>
                  <button onClick={() => retryChange(change.id)}>
                    Retry
                  </button>
                </div>
              )}
            </div>
          ))}
        </div>
      )}
    </div>
  );
}

Other Variations

The implementation above syncs the field_changes table bidirectionally, giving you full visibility into change status on the client. But there are two simpler approaches that reduce overhead when you don’t need complete status tracking:

Insert-Only (Fire and Forget)

For scenarios where you just need to record changes without tracking their status. For example, logging analytics events or recording simple increment/decrement operations. How it works:

Mark the table as insertOnly: true in your client schema
Don’t include the field_changes table in your Sync Rules
Changes are uploaded to the server but never downloaded back to clients

Client schema:

const fieldChanges = new Table(
  {
    table_name: column.text,
    row_id: column.text,
    field_name: column.text,
    new_value: column.text,
    user_id: column.text
  },
  {
    insertOnly: true  // Only allows INSERT operations
  }
);

When to use: Analytics logging, audit trails that don’t need client visibility, simple increment/decrement where conflicts are rare. Tradeoff: No status visibility on the client. You can’t show pending/failed states or implement retry logic.

Pending-Only (Temporary Tracking)

For scenarios where you want to show sync status temporarily but don’t need a permanent history on the client. How it works:

Use a normal table on the client (not insertOnly)
Don’t include the field_changes table in your Sync Rules
Pending changes stay on the client until they’re uploaded and the server processes them
Once the server processes a change and PowerSync syncs the next checkpoint, the change automatically disappears from the client

Client schema:

const pendingChanges = new Table({
  table_name: column.text,
  row_id: column.text,
  field_name: column.text,
  new_value: column.text,
  status: column.text,
  user_id: column.text
});

Show pending indicator:

function SyncIndicator({ taskId }: { taskId: string }) {
  const { data: pending } = useQuery(
    `SELECT COUNT(*) as count FROM pending_changes 
     WHERE row_id = ? AND status = 'pending'`,
    [taskId]
  );
  
  if (!pending?.[0]?.count) return null;
  
  return (
    <div className="sync-badge">
      ⏳ {pending[0].count} change{pending[0].count > 1 ? 's' : ''} syncing...
    </div>
  );
}

When to use: Showing “syncing…” indicators, temporary status tracking without long-term storage overhead, cases where you want automatic cleanup after sync. Tradeoff: Can’t show detailed server-side error messages (unless the server writes to a separate errors table that is in Sync Rules). No long-term history on the client.

Strategy 7: Cumulative Operations (Inventory)

For scenarios like inventory management, simply replacing values causes data loss. When two clerks simultaneously sell the same item while offline, both sales must be honored. The solution is to treat certain fields as deltas rather than absolute values, you subtract incoming quantities from the current stock rather than replacing the count. This requires your backend to recognize which operations should be cumulative. For inventory quantity changes, you apply the delta (e.g., -3 units) to the current value rather than setting it directly. This ensures all concurrent sales are properly recorded without overwriting each other.

Database Schema

Source database (Postgres):

CREATE TABLE inventory (
  id UUID PRIMARY KEY,
  product_id UUID NOT NULL,
  quantity INTEGER NOT NULL DEFAULT 0,
  last_updated TIMESTAMPTZ DEFAULT NOW()
);

-- Prevent negative inventory
ALTER TABLE inventory ADD CONSTRAINT positive_quantity CHECK (quantity >= 0);

Backend: Delta Detection and Application

The key is detecting when an operation should be treated as a delta versus an absolute value. You can identify this through table/field combinations, metadata flags, or operation patterns. Backend API (Node.js):

async function handleInventoryOperation(db, operation) {
  const { table, id, op, opData } = operation;
  
  // Identify cumulative fields for specific tables
  if (table === 'inventory' && 'quantity' in opData) {
    return await applyInventoryDelta(db, operation);
  }
  
  // Default handling for other fields/tables
  return await handleGenericOperation(db, operation);
}

async function applyInventoryDelta(db, operation) {
  const { id, opData } = operation;
  const quantityChange = opData.quantity; // This is the delta, not absolute value
  
  // Get current inventory
  const result = await db.query(
    'SELECT quantity FROM inventory WHERE id = $1',
    [id]
  );
  
  if (!result.rows[0]) {
    return { 
      conflict: 'inventory_not_found',
      message: `Inventory ${id} does not exist`
    };
  }
  
  const currentQuantity = result.rows[0].quantity;
  const newQuantity = currentQuantity + quantityChange;
  
  // Validate: prevent negative inventory
  if (newQuantity < 0) {
    console.warn(`Insufficient stock: ${id} has ${currentQuantity}, attempted change: ${quantityChange}`);
    
    return {
      conflict: 'insufficient_stock',
      message: `Cannot reduce inventory by ${Math.abs(quantityChange)}. Only ${currentQuantity} available.`,
      currentQuantity
    };
  }
  
  // Apply the delta atomically
  await db.query(
    `UPDATE inventory 
     SET quantity = quantity + $1, 
         last_updated = NOW()
     WHERE id = $2`,
    [quantityChange, id]
  );
  
  return { 
    success: true,
    newQuantity,
    previousQuantity: currentQuantity
  };
}

Client Implementation

On the client side, you need to ensure updates are sent as deltas, not absolute values. When a sale occurs, send the change amount: Client code:

// When selling 3 units
await powerSync.execute(
  'UPDATE inventory SET quantity = quantity - ? WHERE id = ?',
  [3, inventoryId] // Send -3 as the delta
);

The backend receives this as a PATCH operation where opData.quantity = -3, which it then adds to the current quantity rather than replacing it.

Alternative Approaches

1. Metadata Flags: Include operation type in metadata to signal delta operations:

await powerSync.execute(
  'UPDATE inventory SET quantity = ?, _metadata = ? WHERE id = ?',
  [
    -3,
    JSON.stringify({ operation_type: 'delta' }),
    inventoryId
  ]
);

Backend checks metadata and applies accordingly. 2. Separate Transactions Table: Track each quantity change as its own row, then aggregate them. This provides full audit history but requires syncing an additional table. 3. Operation-Based Detection: Infer cumulative operations from the pattern. Negative values likely indicate sales (deltas), while large positive values might be absolute restocks requiring different handling.

Using Custom Metadata

Track additional context about operations using the _metadata column.

Enable in Schema

Client schema:

const tasks = new Table(
  {
    title: column.text,
    status: column.text,
  },
  {
    trackMetadata: true  // Enables _metadata column
  }
);

Write Metadata

Client code:

await powerSync.execute(
  'UPDATE tasks SET title = ?, _metadata = ? WHERE id = ?',
  [
    'New title',
    JSON.stringify({ 
      source: 'mobile_app',
      device: 'iPhone 12',
      priority: 'high',
      reason: 'customer_request'
    }),
    taskId
  ]
);

Access in Backend

Backend API (Node.js):

async function processOperation(operation) {
  const metadata = operation.metadata ? JSON.parse(operation.metadata) : {};
  
  // Route high-priority operations differently
  if (metadata.priority === 'high') {
    await processHighPriority(operation);
    return;
  }
  
  // Track which device made the change
  console.log(`Change from: ${metadata.device || 'unknown'}`);
  
  // Custom conflict resolution based on metadata
  if (metadata.reason === 'customer_request') {
    // Customer requests might override other updates
    await forceApplyOperation(operation);
  } else {
    await standardProcessing(operation);
  }
}

Common use cases:

Track which device/app version made the change
Flag operations requiring special handling
Store user context (role, department)
Implement source-based conflict resolution (mobile trumps web)
Pass approval flags or business context

Complete Backend Example

Here’s how to tie it all together in a Node.js backend with Postgres. Backend API (Node.js + Express):

import express from 'express';
import { Pool } from 'pg';

const app = express();
const pool = new Pool({
  connectionString: process.env.DATABASE_URL
});

app.post('/api/data', async (req, res) => {
  const { batch } = req.body;
  const userId = req.user.id; // From auth middleware
  
  const db = await pool.connect();
  
  try {
    await db.query('BEGIN');
    
    for (const operation of batch) {
      // Choose strategy based on table
      if (operation.table === 'orders') {
        await handleOrderOperation(db, operation, userId);
      } else if (operation.table === 'tasks') {
        await handleTaskOperation(db, operation, userId);
      } else {
        // Default handling
        await handleGenericOperation(db, operation);
      }
    }
    
    await db.query('COMMIT');
    res.json({ success: true });
    
  } catch (error) {
    await db.query('ROLLBACK');
    console.error('Operation failed:', error);
    res.status(500).json({ error: error.message });
  } finally {
    db.release();
  }
});

async function handleOrderOperation(db, op, userId) {
  if (op.op === 'PUT') {
    // Use business rule validation (Strategy 4)
    const result = await validateOrderUpdate(db, op);
    if (result.conflict) {
      throw new Error(result.message);
    }
  } else if (op.op === 'PATCH') {
    await handleOrderPatch(db, op, userId);
  } else if (op.op === 'DELETE') {
    await handleOrderDelete(db, op);
  }
}

async function handleTaskOperation(db, op, userId) {
  if (op.op === 'PUT' || op.op === 'PATCH') {
    // Use timestamp detection with conflict recording (Strategy 1 + 5)
    const result = await handleUpdateWithConflictRecording(db, op, userId);
    if (result.conflict && result.conflict !== 'recorded') {
      console.warn('Conflict detected:', result);
    }
  } else if (op.op === 'DELETE') {
    await db.query('DELETE FROM tasks WHERE id = $1', [op.id]);
  }
}

async function handleGenericOperation(db, op) {
  // Default last-write-wins
  if (op.op === 'PUT') {
    const fields = Object.keys(op.opData);
    const values = Object.values(op.opData);
    const placeholders = fields.map((_, i) => `$${i + 1}`).join(', ');
    const updates = fields.map((f, i) => `${f} = $${i + 1}`).join(', ');
    
    await db.query(
      `INSERT INTO ${op.table} (id, ${fields.join(', ')})
       VALUES ($${fields.length + 1}, ${placeholders})
       ON CONFLICT (id) DO UPDATE SET ${updates}`,
      [...values, op.id]
    );
  } else if (op.op === 'PATCH') {
    const fields = Object.keys(op.opData);
    const values = Object.values(op.opData);
    const updates = fields.map((f, i) => `${f} = $${i + 1}`).join(', ');
    
    await db.query(
      `UPDATE ${op.table} SET ${updates} WHERE id = $${fields.length + 1}`,
      [...values, op.id]
    );
  } else if (op.op === 'DELETE') {
    await db.query(`DELETE FROM ${op.table} WHERE id = $1`, [op.id]);
  }
}

app.listen(3000, () => {
  console.log('Backend listening on port 3000');
});

Best Practices

Design for idempotency: Operations arrive multiple times. Check for existing records before inserting, use upserts, or track operation IDs to skip duplicates. Test offline scenarios: Simulate two clients going offline, making conflicting changes, then syncing. Does your resolution strategy behave as expected? Provide clear UI feedback: Show sync status prominently. Users should know when their changes are pending, synced, or conflicted. Consider partial failures: If batch processing fails midway, how do you recover? Use database transactions and mark progress carefully. Log conflicts for analysis: Track how often conflicts occur and why. This data helps you improve UX or adjust resolution strategies. Leverage CRDTs for collaborative docs: For scenarios with real-time collaboration, consider CRDTs to automatically handle concurrent edits. For information on CRDTs, see our separate guide. Collaborative editing without using CRDTs: You can use PowerSync for collaborative text editing without the complexity of CRDTs. See Matthew Weidner’s blog post on collaborative text editing using PowerSync.

​When You Might Need Custom Conflict Resolution

​How Data Flows Through PowerSync

​Client to Backend

​Backend to Client

​Understanding Operations & CrudEntry

​CrudEntry Structure

​What Your Backend Receives

​Implementation Examples

​Strategy 1: Timestamp-Based Detection

​Database Schema

​Backend Conflict Detection

​Strategy 2: Sequence Number Versioning

​Database Schema

​Backend Conflict Detection

​Strategy 3: Field-Level Last Write Wins

​Database Schema

​Client Schema with Metadata

​Client Updates with Timestamps

​Backend Field-Level Resolution

​Strategy 4: Business Rule Validation

​Backend with Business Rules

​Strategy 5: Server-Side Conflict Recording

​Step 1: Create Conflicts Table

​Step 2: Sync Conflicts to Clients

​Step 3: Record Conflicts in Backend

​Step 4: Build Resolution UI

​Strategy 6: Change-Level Status Tracking

​Step 1: Create Change Log Table

​Step 2: Client Writes to Both Tables

​Step 3: Backend Processes Changes

​Step 4: Display Change Status

​Other Variations

​Insert-Only (Fire and Forget)

​Pending-Only (Temporary Tracking)

​Strategy 7: Cumulative Operations (Inventory)

​Database Schema

​Backend: Delta Detection and Application

​Client Implementation

​Alternative Approaches

​Using Custom Metadata

​Enable in Schema

​Write Metadata

​Access in Backend

​Complete Backend Example

​Best Practices

When You Might Need Custom Conflict Resolution

How Data Flows Through PowerSync

Client to Backend

Backend to Client

Understanding Operations & `CrudEntry`

`CrudEntry` Structure

What Your Backend Receives

Implementation Examples

Strategy 1: Timestamp-Based Detection

Database Schema

Backend Conflict Detection

Strategy 2: Sequence Number Versioning

Database Schema

Backend Conflict Detection

Strategy 3: Field-Level Last Write Wins

Database Schema

Client Schema with Metadata

Client Updates with Timestamps

Backend Field-Level Resolution

Strategy 4: Business Rule Validation

Backend with Business Rules

Strategy 5: Server-Side Conflict Recording

Step 1: Create Conflicts Table

Step 2: Sync Conflicts to Clients

Step 3: Record Conflicts in Backend

Step 4: Build Resolution UI

Strategy 6: Change-Level Status Tracking

Step 1: Create Change Log Table

Step 2: Client Writes to Both Tables

Step 3: Backend Processes Changes

Step 4: Display Change Status

Other Variations

Insert-Only (Fire and Forget)

Pending-Only (Temporary Tracking)

Strategy 7: Cumulative Operations (Inventory)

Database Schema

Backend: Delta Detection and Application

Client Implementation

Alternative Approaches

Using Custom Metadata

Enable in Schema

Write Metadata

Access in Backend

Complete Backend Example

Best Practices