Storage API

The Storage API allows Durable Objects to access transactional and strongly consistent storage. A Durable Object’s attached storage is private to its unique instance and cannot be accessed by other objects.

Durable Objects gain access to a persistent Storage API via ctx.storage, on the ctx parameter passed to the Durable Object constructor.

While access to a Durable Object is single-threaded, request executions can still interleave with each other when they wait on I/O, such as when waiting on the promises returned by persistent storage methods or fetch() requests.

The following code snippet shows you how to store and retrieve data using the Storage API.

export class Counter {
  constructor(ctx, env) {
    this.ctx = ctx;
  }

  async fetch(request) {
    let url = new URL(request.url);

    // retrieve data
    let value = (await this.ctx.storage.get("value")) || 0;

    // increment counter and get a new value
    value += 1;

    // store data
    await this.ctx.storage.put("value", value);

    return new Response(value);
  }
}

Methods

SQLite in Durable Objects Beta

The new beta version of Durable Objects is available where each Durable Object has a private, embedded SQLite database. When deploying a new Durable Object class, users can opt-in to a SQLite storage backend in order to access new SQL API. Otherwise, a Durable Object class has a key-value storage backend.

The Storage API comes with several methods, including key-value (KV) API, SQL API, and point-in-time-recovery (PITR) API.

• Durable Object classes with the default, key-value storage backend can use KV API.
• Durable Object classes with the SQLite storage backend can use KV API, SQL API, and PITR API. KV API methods like get(), put(), delete(), or list() store data in a hidden SQLite table.

Each method is implicitly wrapped inside a transaction, such that its results are atomic and isolated from all other storage operations, even when accessing multiple key-value pairs.

get

• get(key string , options Object optional): Promise<any>

  • Retrieves the value associated with the given key. The type of the returned value will be whatever was previously written for the key, or undefined if the key does not exist.

• get(keys Array<string> , options Object optional): Promise<Map<string, any>>

  • Retrieves the values associated with each of the provided keys. The type of each returned value in the Map ↗ will be whatever was previously written for the corresponding key. Results in the Map will be sorted in increasing order of their UTF-8 encodings, with any requested keys that do not exist being omitted. Supports up to 128 keys at a time.

Supported options

• allowConcurrency: boolean

  • By default, the system will pause delivery of I/O events to the Object while a storage operation is in progress, in order to avoid unexpected race conditions. Pass allowConcurrency: true to opt out of this behavior and allow concurrent events to be delivered.

• noCache: boolean

  • If true, then the key/value will not be inserted into the in-memory cache. If the key is already in the cache, the cached value will be returned, but its last-used time will not be updated. Use this when you expect this key will not be used again in the near future. This flag is only a hint. This flag will never change the semantics of your code, but it may affect performance.

put

• put(key string , value any , options Object optional): Promise

  • Stores the value and associates it with the given key. The value can be any type supported by the structured clone algorithm ↗, which is true of most types. Keys are limited to a max size of 2,048 bytes and values are limited to 128 KiB (131,072 bytes).
    
    

• put(entries Object , options Object optional): Promise

  • Takes an Object and stores each of its keys and values to storage.
  • Each value can be any type supported by the structured clone algorithm ↗, which is true of most types.
  • Supports up to 128 key-value pairs at a time. Each key is limited to a maximum size of 2,048 bytes and each value is limited to 128 KiB (131,072 bytes).

delete

• delete(key string , options Object optional): Promise<boolean>

  • Deletes the key and associated value. Returns true if the key existed or false if it did not.

• delete(keys Array<string> , options Object optional): Promise<number>

  • Deletes the provided keys and their associated values. Supports up to 128 keys at a time. Returns a count of the number of key-value pairs deleted.

deleteAll

• deleteAll(options Object optional): Promise

  • Deletes all stored data, effectively deallocating all storage used by the Durable Object. For Durable Objects with a key-value storage backend, deleteAll() removes all keys and associated values for an individual Durable Object. For Durable Objects with a SQLite storage backend, deleteAll() removes the entire contents of a Durable Object’s private SQLite database, including both SQL data and key-value data.
  • For Durable Objects with a key-value storage backend, an in-progress deleteAll() operation can fail, which may leave a subset of data undeleted. Durable Objects with a SQLite storage backend do not have a partial deleteAll() issue because deleteAll() operations are atomic (all or nothing).
  • deleteAll() does not proactively delete Alarms. Use deleteAlarm() to delete an alarm.

Supported options

• put(), delete() and deleteAll() support the following options:

• allowUnconfirmed boolean

  • By default, the system will pause outgoing network messages from the Durable Object until all previous writes have been confirmed flushed to disk. If the write fails, the system will reset the Object, discard all outgoing messages, and respond to any clients with errors instead.

  • This way, Durable Objects can continue executing in parallel with a write operation, without having to worry about prematurely confirming writes, because it is impossible for any external party to observe the Object’s actions unless the write actually succeeds.

  • After any write, subsequent network messages may be slightly delayed. Some applications may consider it acceptable to communicate on the basis of unconfirmed writes. Some programs may prefer to allow network traffic immediately. In this case, set allowUnconfirmed to true to opt out of the default behavior.

  • If you want to allow some outgoing network messages to proceed immediately but not others, you can use the allowUnconfirmed option to avoid blocking the messages that you want to proceed and then separately call the sync() method, which returns a promise that only resolves once all previous writes have successfully been persisted to disk.

• noCache boolean

  • If true, then the key/value will be discarded from memory as soon as it has completed writing to disk.

  • Use noCache if the key will not be used again in the near future. noCache will never change the semantics of your code, but it may affect performance.

  • If you use get() to retrieve the key before the write has completed, the copy from the write buffer will be returned, thus ensuring consistency with the latest call to put().

Automatic write coalescing

If you invoke put() (or delete()) multiple times without performing any await in the meantime, the operations will automatically be combined and submitted atomically. In case of a machine failure, either all of the writes will have been stored to disk or none of the writes will have been stored to disk.

Write buffer behavior

The put() method returns a Promise, but most applications can discard this promise without using await. The Promise usually completes immediately, because put() writes to an in-memory write buffer that is flushed to disk asynchronously. However, if an application performs a large number of put() without waiting for any I/O, the write buffer could theoretically grow large enough to cause the isolate to exceed its 128 MB memory limit. To avoid this scenario, such applications should use await on the Promise returned by put(). The system will then apply backpressure onto the application, slowing it down so that the write buffer has time to flush. Using await will disable automatic write coalescing.

list

• list(options Object optional): Promise<Map<string, any>>

  • Returns all keys and values associated with the current Durable Object in ascending sorted order based on the keys’ UTF-8 encodings.

  • The type of each returned value in the Map ↗ will be whatever was previously written for the corresponding key.

  • Be aware of how much data may be stored in your Durable Object before calling this version of list without options because all the data will be loaded into the Durable Object’s memory, potentially hitting its limit. If that is a concern, pass options to list as documented below.

Supported options

• start string

  • Key at which the list results should start, inclusive.

• startAfter string

  • Key after which the list results should start, exclusive. Cannot be used simultaneously with start.

• end string

  • Key at which the list results should end, exclusive.

• prefix string

  • Restricts results to only include key-value pairs whose keys begin with the prefix.

• reverse boolean

  • If true, return results in descending order instead of the default ascending order.
  • Enabling reverse does not change the meaning of start, startKey, or endKey. start still defines the smallest key in lexicographic order that can be returned (inclusive), effectively serving as the endpoint for a reverse-order list. end still defines the largest key in lexicographic order that the list should consider (exclusive), effectively serving as the starting point for a reverse-order list.

• limit number

  • Maximum number of key-value pairs to return.

• allowConcurrency boolean

  • Same as the option to get(), above.

• noCache boolean

  • Same as the option to get(), above.

transaction

• transaction(closureFunction(txn)): Promise

  • Runs the sequence of storage operations called on txn in a single transaction that either commits successfully or aborts.

  • Explicit transactions are no longer necessary. Any series of write operations with no intervening await will automatically be submitted atomically, and the system will prevent concurrent events from executing while await a read operation (unless you use allowConcurrency: true). Therefore, a series of reads followed by a series of writes (with no other intervening I/O) are automatically atomic and behave like a transaction.

• txn

  • Provides access to the put(), get(), delete() and list() methods documented above to run in the current transaction context. In order to get transactional behavior within a transaction closure, you must call the methods on the txn Object instead of on the top-level ctx.storage Object.
    
    Also supports a rollback() function that ensures any changes made during the transaction will be rolled back rather than committed. After rollback() is called, any subsequent operations on the txn Object will fail with an exception. rollback() takes no parameters and returns nothing to the caller.

  • When using the SQLite-backed storage engine, the txn object is obsolete. Any storage operations performed directly on the ctx.storage object, including SQL queries using ctx.storage.sql.exec(), will be considered part of the transaction.

transactionSync

• transactionSync(callback): any

  • Only available when using the SQLite-backed storage engine.

  • Invokes callback() wrapped in a transaction, and returns its result.

  • If callback() throws an exception, the transaction will be rolled back.

  • The callback must complete synchronously, that is, it should not be declared async nor otherwise return a Promise. Only synchronous storage operations can be part of the transaction. This is intended for use with SQL queries using ctx.storage.sql.exec(), which complete sychronously.

sync

• sync() : Promise

  • Synchronizes any pending writes to disk.

  • This is similar to normal behavior from automatic write coalescing. If there are any pending writes in the write buffer (including those submitted with the allowUnconfirmed option), the returned promise will resolve when they complete. If there are no pending writes, the returned promise will be already resolved.

getAlarm

• getAlarm(options Object optional): Promise<Number | null>

  • Retrieves the current alarm time (if set) as integer milliseconds since epoch. The alarm is considered to be set if it has not started, or if it has failed and any retry has not begun. If no alarm is set, getAlarm() returns null.

Supported options

• Same options as get(), but without noCache.

setAlarm

• setAlarm(scheduledTime Date | number , options Object optional): Promise

  • Sets the current alarm time, accepting either a JavaScript Date, or integer milliseconds since epoch.

    
    If setAlarm() is called with a time equal to or before Date.now(), the alarm will be scheduled for asynchronous execution in the immediate future. If the alarm handler is currently executing in this case, it will not be canceled. Alarms can be set to millisecond granularity and will usually execute within a few milliseconds after the set time, but can be delayed by up to a minute due to maintenance or failures while failover takes place.

deleteAlarm

• deleteAlarm(options Object optional): Promise

  • Deletes the alarm if one exists. Does not cancel the alarm handler if it is currently executing.

Supported options

• setAlarm() and deleteAlarm() support the same options as put(), but without noCache.

sql.exec

SQLite in Durable Objects Beta

SQL API methods accessed with ctx.storage.sql are only allowed on Durable Object classes with SQLite storage backend and will return an error if called on Durable Object classes with a key-value storage backend.

ctx.storage.sql.exec(query: string , …bindings: any[] ): SqlStorageCursor

Parameters

• query: string
  • The SQL query string to be executed. query can contain ? placeholders for parameter bindings. Multiple SQL statements, separated with a semicolon, can be executed in the query. With multiple SQL statements, any parameter bindings are applied to the last SQL statement in the query, and the returned cursor is only for the last SQL statement.
• bindings: any[] Optional
  • Optional variable number of arguments that correspond to the ? placeholders in query.

Returns

A cursor (SqlStorageCursor) to iterate over query row results as objects. SqlStorageCursor is a JavaScript Iterable ↗, which supports iteration using for (let row of cursor). SqlStorageCursor is also a JavaScript Iterator ↗, which supports iteration using cursor.next().

SqlStorageCursor supports the following methods:

• next()
  • Returns an object representing the next value of the cursor. The returned object has done and value properties adhering to the JavaScript Iterator ↗. done is set to false when a next value is present, and value is set to the next row object in the query result. done is set to true when the entire cursor is consumed, and no value is set.
• toArray()
  • Iterates through remaining cursor value(s) and returns an array of returned row objects.
• one()
  • Returns a row object if query result has exactly one row. If query result has zero rows or more than one row, one() throws an exception.
• raw(): Iterator
  • Returns an Iterator over the same query results, with each row as an array of column values (with no column names) rather than an object.
  • Returned Iterator supports next(), toArray(), and one() methods above.
  • Returned cursor and raw() iterator iterate over the same query results and can be combined. For example:

  let cursor = this.sql.exec("SELECT * FROM artist ORDER BY artistname ASC;");
  let rawResult = cursor.raw().next();

  if (!rawResult.done) {
    console.log(rawResult.value); // prints [ 123, 'Alice' ]
  } else {
    // query returned zero results
  }

  console.log(cursor.toArray()); // prints [{ artistid: 456, artistname: 'Bob' },{ artistid: 789, artistname: 'Charlie' }]

SqlStorageCursor had the following properties:

• columnNames: string[]
  • The column names of the query in the order they appear in each row array returned by the raw iterator.
• rowsRead: number
• The number of rows read so far as part of this SQL query. This may increase as you iterate the cursor. The final value is used for SQL billing.
• rowsWritten: number
• The number of rows written so far as part of this SQL query. This may increase as you iterate the cursor. The final value is used for SQL billing.

Note that sql.exec() cannot execute transaction-related statements like BEGIN TRANSACTION or SAVEPOINT. Instead, use the ctx.storage.transaction() or ctx.storage.transactionSync() APIs to start a transaction, and then execute SQL queries in your callback.

Examples

SQL API examples below use the following SQL schema:

import { DurableObject } from "cloudflare:workers";

export class MyDurableObject extends DurableObject {
  sql: SqlStorage
  constructor(ctx: DurableObjectState, env: Env) {
    super(ctx, env);
    this.sql = ctx.storage.sql;

    this.sql.exec(`CREATE TABLE IF NOT EXISTS artist(
      artistid    INTEGER PRIMARY KEY,
      artistname  TEXT
    );INSERT INTO artist (artistid, artistname) VALUES
      (123, 'Alice'),
      (456, 'Bob'),
      (789, 'Charlie');`
    );
  }
}

Iterate over query results as row objects:

  let cursor = this.sql.exec("SELECT * FROM artist;");

  for (let row of cursor) {
    // Iterate over row object and do something
  }

Convert query results to an array of row objects:

  // Return array of row objects: [{"artistid":123,"artistname":"Alice"},{"artistid":456,"artistname":"Bob"},{"artistid":789,"artistname":"Charlie"}]
  let resultsArray1 = this.sql.exec("SELECT * FROM artist;").toArray();
  // OR
  let resultsArray2 = Array.from(this.sql.exec("SELECT * FROM artist;"));
  // OR
  let resultsArray3 = [...this.sql.exec("SELECT * FROM artist;")]; // JavaScript spread syntax

Convert query results to an array of row values arrays:

  // Returns [[123,"Alice"],[456,"Bob"],[789,"Charlie"]]
  let cursor = this.sql.exec("SELECT * FROM artist;");
  let resultsArray = cursor.raw().toArray();

  // Returns ["artistid","artistname"]
  let columnNameArray = this.sql.exec("SELECT * FROM artist;").columnNames.toArray();

Get first row object of query results:

  // Returns {"artistid":123,"artistname":"Alice"}
  let firstRow = this.sql.exec("SELECT * FROM artist ORDER BY artistname DESC;").toArray()[0];

Check if query results have exactly one row:

  // returns error
  this.sql.exec("SELECT * FROM artist ORDER BY artistname ASC;").one();

  // returns { artistid: 123, artistname: 'Alice' }
  let oneRow = this.sql.exec("SELECT * FROM artist WHERE artistname = ?;", "Alice").one()

Returned cursor behavior:

  let cursor = this.sql.exec("SELECT * FROM artist ORDER BY artistname ASC;");
  let result = cursor.next();
  if (!result.done) {
    console.log(result.value); // prints { artistid: 123, artistname: 'Alice' }
  } else {
    // query returned zero results
  }

  let remainingRows = cursor.toArray();
  console.log(remainingRows); // prints [{ artistid: 456, artistname: 'Bob' },{ artistid: 789, artistname: 'Charlie' }]

Returned cursor and raw() iterator iterate over the same query results:

  let cursor = this.sql.exec("SELECT * FROM artist ORDER BY artistname ASC;");
  let result = cursor.raw().next();

  if (!result.done) {
    console.log(result.value); // prints [ 123, 'Alice' ]
  } else {
    // query returned zero results
  }

  console.log(cursor.toArray()); // prints [{ artistid: 456, artistname: 'Bob' },{ artistid: 789, artistname: 'Charlie' }]

sql.exec().rowsRead():

  let cursor = this.sql.exec("SELECT * FROM artist;");
  cursor.next()
  console.log(cursor.rowsRead); // prints 1

  cursor.toArray(); // consumes remaining cursor
  console.log(cursor.rowsRead); // prints 3

sql.databaseSize

ctx.storage.sql.databaseSize: number

Returns

The current SQLite database size in bytes.

  let size = ctx.storage.sql.databaseSize;

Point in time recovery

For Durable Objects classes with SQL storage, the following point-in-time-recovery (PITR) API methods are available to restore a Durable Object’s embedded SQLite database to any point in time in the past 30 days. These methods apply to the entire SQLite database contents, including both the object’s stored SQL data and stored key-value data using the key-value put() API. The PITR API is not supported in local development because a durable log of data changes is not stored locally.

The PITR API represents points in times using “bookmarks”. A bookmark is a mostly alphanumeric string like 0000007b-0000b26e-00001538-0c3e87bb37b3db5cc52eedb93cd3b96b. Bookmarks are designed to be lexically comparable: a bookmark representing an earlier point in time compares less than one representing a later point, using regular string comparison.

ctx.storage.getCurrentBookmark(): Promise<string>

• Returns a bookmark representing the current point in time in the object’s history.

ctx.storage.getBookmarkForTime(timestamp: number | Date ): Promise<string>

• Returns a bookmark representing approximately the given point in time, which must be within the last 30 days. If the timestamp is represented as a number, it is converted to a date as if using new Date(timestamp).

ctx.storage.onNextSessionRestoreBookmark(bookmark: string ): Promise<string>

• Configure the Durable Object so that the next time it restarts, it should restore its storage to exactly match what the storage contained at the given bookmark. After calling this, the application should typically invoke ctx.abort() to restart the Durable Object, thus completing the point-in-time recovery.

This method returns a special bookmark representing the point in time immediately before the recovery takes place (even though that point in time is still technically in the future). Thus, after the recovery completes, it can be undone by performing a second recovery to this bookmark.

  let now = new Date();
  // restore to 2 days ago
  let bookmark = ctx.storage.getBookmarkForTime(now - 2);
  ctx.storage.onNextSessionRestoreBookmark(bookmark);

Related resources

• Durable Objects: Easy, Fast, Correct – Choose Three ↗.
• WebSockets API.