UPSERT operation in AQL

An UPSERT operation either modifies an existing document, or creates a new document if it does not exist

UPSERT looks up a single document that matches the provided example. If there is no match, an insert operation is executed to create a document. If a document is found, you can either update or replace the document. These subtypes are called upsert (update or insert) and repsert (replace or insert).

Each UPSERT operation is restricted to a single collection, and the collection name must not be dynamic. Only a single UPSERT statement per collection is allowed per AQL query, and it cannot be followed by read or write operations that access the same collection, by traversal operations, or AQL functions that can read documents.


The syntax for an upsert operation:

UPSERT searchExpression
INSERT insertExpression
UPDATE updateExpression
IN collection

The syntax for a repsert operation:

UPSERT searchExpression
INSERT insertExpression
REPLACE updateExpression
IN collection

Both variants can optionally end with an OPTIONS { … } clause.

When using the UPDATE variant of the UPSERT operation, the found document is partially updated, meaning only the attributes specified in updateExpression are updated or added. When using the REPLACE variant of UPSERT (repsert), the found document is replaced with the content of updateExpression.

Updating a document modifies the document’s revision number with a server-generated value. The system attributes _id, _key, and _rev cannot be updated, but _from and _to can be modified.

The searchExpression contains the document to be looked for. It must be an object literal (UPSERT { <key>: <value>, ... } ...) without dynamic attribute names. In case no such document can be found in collection, a new document is inserted into the collection as specified in the insertExpression.

In case at least one document in collection matches the searchExpression, it is updated using the updateExpression. When more than one document in the collection matches the searchExpression, it is undefined which of the matching documents is updated. It is therefore often sensible to make sure by other means (such as unique indexes, application logic etc.) that at most one document matches searchExpression.

The following query looks for a document in the users collection with a specific name attribute value. If the document exists, its logins attribute is increased by one. If it does not exist, a new document is inserted, consisting of the attributes name, logins, and dateCreated:

UPSERT { name: 'superuser' } 
INSERT { name: 'superuser', logins: 1, dateCreated: DATE_NOW() } 
UPDATE { logins: OLD.logins + 1 } IN users

Note that in the UPDATE case it is possible to refer to the previous version of the document using the OLD pseudo-value.

Query options


The ignoreErrors option can be used to suppress query errors that may occur when trying to violate unique key constraints.


When updating an attribute to the null value, ArangoDB does not remove the attribute from the document but stores this null value. To remove attributes in an update operation, set them to null and set the keepNull option to false. This removes the attributes you specify but not any previously stored attributes with the null value:

UPSERT { _key: "mary" }
INSERT { _key: "mary", name: "Mary", notNeeded: 123 }
UPDATE { foobar: true, notNeeded: null }
IN users OPTIONS { keepNull: false }

If no document with the key mary exists, the above query creates such a user document with a notNeeded attribute. If it exists already, it removes the notNeeded attribute from the document and updates the foobar attribute normally.

Only top-level attributes and sub-attributes can be removed this way (e.g. { attr: { sub: null } }) but not attributes of objects that are nested inside of arrays (e.g. { attr: [ { nested: null } ] }).


The option mergeObjects controls whether object contents are merged if an object attribute is present in both the UPDATE query and in the to-be-updated document.

The default value for mergeObjects is true, so there is no need to specify it explicitly.


To make sure data are durable when an update query returns, there is the waitForSync query option.


In order to not accidentally update documents that have been written and updated since you last fetched them you can use the option ignoreRevs to either let ArangoDB compare the _rev value and only succeed if they still match, or let ArangoDB ignore them (default):

FOR i IN 1..1000
  UPSERT { _key: CONCAT('test', i)}
    INSERT {foobar: false}
    UPDATE {_rev: "1287623", foobar: true }
  IN users OPTIONS { ignoreRevs: false }
You need to add the _rev value in the updateExpression. It is not used within the searchExpression. Even worse, if you use an outdated _rev in the searchExpression, UPSERT triggers the INSERT path instead of the UPDATE path, because it has not found a document exactly matching the searchExpression.


The RocksDB engine does not require collection-level locks. Different write operations on the same collection do not block each other, as long as there are no write-write conflicts on the same documents. From an application development perspective it can be desired to have exclusive write access on collections, to simplify the development. Note that writes do not block reads in RocksDB. Exclusive access can also speed up modification queries, because we avoid conflict checks.

Use the exclusive option to achieve this effect on a per query basis:

FOR i IN 1..1000
  UPSERT { _key: CONCAT('test', i) }
  INSERT { foobar: false }
  UPDATE { foobar: true }
  IN users OPTIONS { exclusive: true }


The indexHint option is used as a hint for the document lookup performed as part of the UPSERT operation, and can help in cases such as UPSERT not picking the best index automatically.

UPSERT { a: 1234 }
  INSERT { a: 1234, name: "AB" }
  UPDATE { name: "ABC" } IN myCollection
  OPTIONS { indexHint: "index_name" }

The index hint is passed through to an internal FOR loop that is used for the lookup. Also see indexHint Option of the FOR Operation.

Inverted indexes cannot be used for UPSERT lookups.


Makes the index or indexes specified in indexHint mandatory if enabled. The default is false. Also see forceIndexHint Option of the FOR Operation.

UPSERT { a: 1234 }
  INSERT { a: 1234, name: "AB" }
  UPDATE { name: "ABC" } IN myCollection
  OPTIONS { indexHint:  , forceIndexHint: true }

Returning documents

UPSERT statements can optionally return data. To do so, they need to be followed by a RETURN statement (intermediate LET statements are allowed, too). These statements can optionally perform calculations and refer to the pseudo-values OLD and NEW. In case the upsert performed an insert operation, OLD has a value of null. In case the upsert performed an update or replace operation, OLD contains the previous version of the document, before update/replace.

NEW is always populated. It contains the inserted document in case the upsert performed an insert, or the updated/replaced document in case it performed an update/replace.

This can also be used to check whether the upsert has performed an insert or an update internally:

UPSERT { name: 'superuser' } 
INSERT { name: 'superuser', logins: 1, dateCreated: DATE_NOW() } 
UPDATE { logins: OLD.logins + 1 } IN users
RETURN { doc: NEW, type: OLD ? 'update' : 'insert' }

Transactionality and Limitations

  • On a single server, upserts are generally executed transactionally in an all-or-nothing fashion.

    For sharded collections in cluster deployments, the entire query and/or upsert operation may not be transactional, especially if it involves different shards, DB-Servers, or both.

  • Queries may execute intermediate transaction commits in case the running transaction (AQL query) hits the specified size thresholds. This writes the data that has been modified so far and it is not rolled back in case of a later abort/rollback of the transaction.

    Such intermediate commits can occur for UPSERT operations over all documents of a large collection, for instance. This has the side-effect that atomicity of this operation cannot be guaranteed anymore and ArangoDB cannot guarantee that “read your own writes” in upserts work.

    This is only an issue if you write a query where your search condition would hit the same document multiple times, and only if you have large transactions. You can adjust the behavior of the RocksDB storage engine by increasing the intermediateCommit thresholds for data size and operation counts.

  • The lookup and the insert/update/replace parts are executed one after another, so that other operations in other threads can happen in between. This means if multiple UPSERT queries run concurrently, they may all determine that the target document does not exist and then create it multiple times!

    Note that due to this gap between the lookup and insert/update/replace, even with a unique index, duplicate key errors or conflicts can occur. But if they occur, the application/client code can execute the same query again.

    To prevent this from happening, you should add a unique index to the lookup attribute(s). Note that in the cluster a unique index can only be created if it is equal to the shard key attribute of the collection or at least contains it as a part.

    An alternative to making an UPSERT statement work atomically is to use the exclusive option to limit write concurrency for this collection to 1, which helps avoiding conflicts but is bad for throughput!

  • UPSERT operations do not observe their own writes correctly in cluster deployments. They only do for OneShard databases with the cluster-one-shard optimizer rule active.

    If upserts in a query create new documents and would then semantically hit the same documents again, the operation may incorrectly use the INSERT branch to create more documents instead of the UPDATE/REPLACE branch to update the previously created documents.

    If upserts find existing documents for updating/replacing, you can access the current document via the OLD pseudo-variable, but this may hold the initial version of the document from before the query even if it has been modified by UPSERT in the meantime.

  • The lookup attribute(s) from the search expression should be indexed in order to improve the UPSERT performance. Ideally, the search expression contains the shard key, as this allows the lookup to be restricted to a single shard.