HTTP request handling in ArangoDB

Protocol

ArangoDB exposes its API via HTTP, making the server accessible easily with a variety of clients and tools (e.g. browsers, curl, telnet). The communication can optionally be SSL-encrypted.

Additionally, there is a custom binary protocol called VelocyStream which can be used for better throughput. HTTP requests are easily mappable to VelocyStream and no separate documentation exists as the API is essentially the same for both network protocols.

ArangoDB uses the standard HTTP methods (e.g. GET, POST, PUT, DELETE) plus the PATCH method described in RFC 5789.

Most server APIs expect clients to send any payload data in JSON format or ArangoDB’s custom VelocyPack binary format. Details on the expected format and JSON attributes can be found in the documentation of the individual API endpoints.

Clients sending requests to ArangoDB must use either of the following protocols:

  • HTTP 1.1
  • HTTP 2
  • VelocyStream

Other HTTP versions or protocols are not supported by ArangoDB.

Clients are required to include the Content-Length HTTP header with the correct content length in every request that can have a body (e.g. POST, PUT or PATCH) request. ArangoDB will not process requests without a Content-Length header - thus chunked transfer encoding for POST-documents is not supported.

HTTP Keep-Alive

ArangoDB supports HTTP keep-alive. If the client does not send a Connection header in its request, ArangoDB will assume the client wants to keep alive the connection. If clients do not wish to use the keep-alive feature, they should explicitly indicate that by sending a Connection: Close HTTP header in the request.

The default Keep-Alive timeout can be specified at server start using the --http.keep-alive-timeout startup option.

Establishing TCP connections is expensive, since it takes several round-trips between the communication parties. Therefore you can use connection keep-alive to send several HTTP request over one TCP-connection; each request is treated independently by definition. You can use this feature to build up a so called connection pool with several established connections in your client application, and dynamically re-use one of those then idle connections for subsequent requests.

Switch protocols

Connections are initialized expecting the HTTP 1.1 protocol by default. To use other protocols the client must indicate this to the server so that the protocol may be switched.

Upgrading to HTTP 2 is supported according to the ways outlined in RFC 7540 Section 3, from non-encrypted connections as well as encrypted connections.

On non-encrypted connections with http scheme in the URI clients may use HTTP 1.1 initially until an upgrade is performed. Upgrading the connection is initiated by sending a request with the Upgrade: h2c header and exactly one HTTP2-Settings header. The server will then respond with 101 Switching Protocols and begin using HTTP/2. See the RFC for details.

For non-encrypted TCP connections ArangoDB also supports Starting HTTP/2 with Prior Knowledge, as specified in RFC 7540 Section 3.4. The server will check the first 24 octets received over the connection and compare it to the HTTP 2 connection preface PRI * HTTP/2.0\r\n\r\nSM\r\n\r\n, as outlined in Section 5.

On TLS encrypted connections with https scheme in the URI ArangoDB supports the ALPN extension with the h2 protocol identifier. This means the connection may switch to using HTTP/2 right away after a successful TLS handshake.

An upgrade to the VelocyStream protocol may happen by sending VST/1.1\r\n\r\n (11 octets) to the server before sending anything else. The server will then start using VelocyStream 1.1. Sending anything else is an error.

Request execution

ArangoDB supports both blocking and non-blocking HTTP requests. Clients can choose the appropriate method on a per-request level based on their throughput, control flow, and durability requirements.

Blocking execution

ArangoDB is a multi-threaded server, allowing the processing of multiple client requests at the same time. Communication handling and the actual work can be performed by multiple worker threads in parallel.

Still, clients need to wait for their requests to be processed by the server, and thus keep one connection of a pool occupied. By default, the server fully processes an incoming request and then returns the result to the client when the operation has finished. The client must wait for the server’s HTTP response before it can send additional requests over the same connection. For clients that are single-threaded, or blocking on I/O themselves, or both, it may not be optimal to wait idle for the full server response.

Furthermore, note that even if the client closes the HTTP connection, the request running on the server still continues until it completes and only then notices that the client no longer listens for the result. Thus closing the connection does not help to abort a long running query.

Non-blocking execution

You can let ArangoDB execute requests asynchronously in two different ways:

  • Submit requests for async execution, without the ability to cancel requests or access the results (“Fire and forget”)

  • Submit requests for async execution and let the server store the results for later retrieval

Fire and forget

To reduce blocking on the client-side, ArangoDB offers a generic mechanism for non-blocking, asynchronous execution: clients can add a x-arango-async: true HTTP header to any of their requests, marking them as to be executed asynchronously on the server.

ArangoDB puts asynchronous requests into an in-memory job queue and instantly returns an HTTP 202 Accepted status code to the client and thus finishes this HTTP request. The server executes the jobs from the queue asynchronously as fast as possible, while clients can continue to do other work.

If the server queue is full (i.e. contains as many jobs as specified by the --server.maximal-queue-size startup option), then the request is rejected instantly with an HTTP 503 Service Unavailable status code.

Asynchronous execution decouples the request/response handling from the actual work to be performed, allowing fast server responses and greatly reducing wait time for clients. Overall this allows for much higher throughput than if clients would always wait for the server’s response.

Keep in mind that the asynchronous execution is just “fire and forget”. Clients get any of their asynchronous requests answered with a generic HTTP 202 Accepted response. At the time the server sends this response, it does not know whether the requested operation can be carried out successfully, as the actual operation execution happens at some later point. Therefore, clients cannot make a decision based on the server response and must rely on their requests being valid and processable by the server.

Additionally, the server’s asynchronous job queue is an in-memory data structure, meaning not-yet processed jobs from the queue are lost in case of a crash. Client will not know whether they were processed or not and should therefore not use the asynchronous feature when they have strict durability requirements or if they rely on the immediate result of the request they send.

Finally, note that it is not possible to cancel such a fire and forget job, since you don’t get any handle to identify it later on. If you need to cancel requests, use Async execution and later result retrieval.

Async execution and later result retrieval

By adding a x-arango-async: store HTTP header to a request, clients can instruct the ArangoDB server to execute the operation asynchronously but also store the operation result in memory for a later retrieval. The server returns a job ID in the x-arango-async-id HTTP response header. The client can use this ID in conjunction with the /_api/job endpoint to access the result. See the HTTP interface for jobs for details.

Clients can ask the ArangoDB server via the async jobs API which results are ready for retrieval, and which are not. Clients can also use the async jobs API to retrieve the original results of an already executed async job by passing it the originally returned job ID. The server then returns the job result as if the job was executed normally. Furthermore, clients can cancel running async jobs by their job ID.

ArangoDB keeps all results of jobs initiated with the x-arango-async: store header. Results are removed from the server only if a client explicitly asks the server for a specific result.

The async jobs API also provides methods for garbage collection that clients can use to get rid of “old”, not fetched results. Clients should call this method periodically because ArangoDB does not artificially limit the number of not-yet-fetched results.

It is thus a client responsibility to store only as many results as needed and to fetch available results as soon as possible, or at least to clean up not fetched results from time to time.

The job queue and the results are kept in-memory only on the server, which means they are lost in case of a crash.

Canceling asynchronous jobs

A running async query can internally be executed by C++ code or by JavaScript code. For example, CRUD operations are executed directly in C++, whereas AQL queries and transactions may be executed by JavaScript code, depending on the AQL functions and the transaction type you use. The job cancelation only works for JavaScript code, since the mechanism used is simply to trigger an uncatchable exception in the JavaScript thread, which is caught on the C++ level, which in turn leads to the cancelation of the job. No result can be retrieved later because all data about the request is discarded.

If you cancel a job running on a Coordinator of a cluster, then only the code running on the Coordinator is stopped. There may remain tasks within the cluster which have already been distributed to the DB-Servers and it is not possible to cancel them as well.

Async execution and authentication

If a request requires authentication, the authentication procedure is run before queueing. The request is only queued if the authentication is successful. If the Otherwise, it is not queued but rejected instantly in the same way as a regular, non-queued request.

Error handling

The following should be noted about how ArangoDB handles client errors in its HTTP layer:

  • client requests using an HTTP version signature different than HTTP/1.0 or HTTP/1.1 will get an HTTP 505 (HTTP Version Not Supported) error in return.
  • ArangoDB will reject client requests with a negative value in the Content-Length request header by closing the connection.
  • ArangoDB doesn’t support POST with Transfer-Encoding: chunked which forbids the Content-Length header above.
  • the maximum URL length accepted by ArangoDB is 16K. Incoming requests with longer URLs will be rejected with an HTTP 414 (Request-URI too long) error.
  • if the client sends a Content-Length header with a value bigger than 0 for an HTTP GET, HEAD, or DELETE request, ArangoDB will process the request, but will write a warning to its log file.
  • when the client sends a Content-Length header that has a value that is lower than the actual size of the body sent, ArangoDB will respond with HTTP 400 (Bad Request).
  • if clients send a Content-Length value bigger than the actual size of the body of the request, ArangoDB will wait for about 90 seconds for the client to complete its request. If the client does not send the remaining body data within this time, ArangoDB will close the connection. Clients should avoid sending such malformed requests as this will block one TCP connection, and may lead to a temporary file descriptor leak.
  • when clients send a body or a Content-Length value bigger than the maximum allowed value (1 GB), ArangoDB will respond with HTTP 413 (Payload Too Large).
  • if the overall length of the HTTP headers a client sends for one request exceeds the maximum allowed size (1 MB), the server will fail with HTTP 431 (Request Header Fields Too Large).
  • if clients request an HTTP method that is not supported by the server, ArangoDB will return with HTTP 405 (Method Not Allowed). ArangoDB offers general support for the following HTTP methods:
    • GET
    • POST
    • PUT
    • DELETE
    • HEAD
    • PATCH
    • OPTIONS

    Please note that not all server actions allow using all of these HTTP methods. You should look up the supported methods for each method you intend to use in the manual.

    Requests using any other HTTP method (such as for example CONNECT, TRACE etc.) will be rejected by ArangoDB as mentioned before.

  • if the backend is temporarily unavailable, the server will return HTTP 503 (Service Unavailable). Common circumstances are:
    • during server start or shutdown, when the network port is open but the HTTP service is not available
    • when the queue is full
    • when a Coordinator cannot reach a DB-Server

    Clients may retry requests but they might not be idempotent.

Cross-Origin Resource Sharing (CORS) requests

ArangoDB automatically handles CORS requests as described below.

Preflight

When a browser is told to make a cross-origin request that includes explicit headers, credentials or uses HTTP methods other than GET or POST, it will first perform a so-called preflight request using the OPTIONS method.

ArangoDB will respond to OPTIONS requests with an HTTP 200 status response with an empty body. Since preflight requests are not expected to include or even indicate the presence of authentication credentials even when they will be present in the actual request, ArangoDB does not enforce authentication for OPTIONS requests even when authentication is enabled.

ArangoDB will set the following headers in the response:

  • access-control-allow-credentials: will be set to false by default. For details on when it will be set to true see the next section on cookies.

  • access-control-allow-headers: will be set to the exact value of the request’s access-control-request-headers header or omitted if no such header was sent in the request.

  • access-control-allow-methods: will be set to a list of all supported HTTP headers regardless of the target endpoint. In other words that a method is listed in this header does not guarantee that it will be supported by the endpoint in the actual request.

  • access-control-allow-origin: will be set to the exact value of the request’s origin header.

  • access-control-expose-headers: will be set to a list of response headers used by the ArangoDB HTTP API.

  • access-control-max-age: will be set to an implementation-specific value.

Actual request

If a request using any other HTTP method than OPTIONS includes an origin header, ArangoDB will add the following headers to the response:

  • access-control-allow-credentials: will be set to false by default. For details on when it will be set to true see the next section on cookies.

  • access-control-allow-origin: will be set to the exact value of the request’s origin header.

  • access-control-expose-headers: will be set to a list of response headers used by the ArangoDB HTTP API.

When making CORS requests to endpoints of Foxx services, the value of the access-control-expose-headers header will instead be set to a list of response headers used in the response itself (but not including the access-control- headers). Note that Foxx services may override this behavior.

Cookies and authentication

In order for the client to be allowed to correctly provide authentication credentials or handle cookies, ArangoDB needs to set the access-control-allow-credentials response header to true instead of false.

ArangoDB will automatically set this header to true if the value of the request’s origin header matches a trusted origin in the http.trusted-origin configuration option. To make ArangoDB trust a certain origin, you can provide a startup option when running arangod like this:

--http.trusted-origin "http://localhost:8529"

To specify multiple trusted origins, the option can be specified multiple times. Alternatively you can use the special value "*" to trust any origin:

--http.trusted-origin "*"

Note that browsers will not actually include credentials or cookies in cross-origin requests unless explicitly told to do so:

  • When using the Fetch API you need to set the credentials option to include.

    fetch("./", { credentials:"include" }).then(/* … */)
    
  • When using XMLHttpRequest you need to set the withCredentials option to true.

    var xhr = new XMLHttpRequest();
    xhr.open('GET', 'https://example.com/', true);
    xhr.withCredentials = true;
    xhr.send(null);
    
  • When using jQuery you need to set the xhrFields option:

    $.ajax({
       url: 'https://example.com',
       xhrFields: {
          withCredentials: true
       }
    });
    

HTTP method overriding

HTTP method overriding is deprecated from version 3.9.0 on and should no longer be used.

ArangoDB provides a startup option --http.allow-method-override. This option can be set to allow overriding the HTTP request method (e.g. GET, POST, PUT, DELETE, PATCH) of a request using one of the following custom HTTP headers:

  • x-http-method-override
  • x-http-method
  • x-method-override

This allows using HTTP clients that do not support all “common” HTTP methods such as PUT, PATCH and DELETE. It also allows bypassing proxies and tools that would otherwise just let certain types of requests (e.g. GET and POST) pass through.

Enabling this option may impose a security risk, so it should only be used in very controlled environments. Thus the default value for this option is false (no method overriding allowed). You need to enable it explicitly if you want to use this feature.

Load-balancer support

When running in cluster mode, ArangoDB exposes some APIs which store request state data on specific Coordinator nodes, and thus subsequent requests which require access to this state must be served by the Coordinator node which owns this state data. In order to support function behind a load-balancer, ArangoDB can transparently forward requests within the cluster to the correct node. If a request is forwarded, the response will contain the following custom HTTP header whose value will be the ID of the node which actually answered the request:

  • x-arango-request-forwarded-to

The following APIs may use request forwarding:

  • /_api/control_pregel
  • /_api/cursor
  • /_api/job
  • /_api/replication
  • /_api/query
  • /_api/tasks
  • /_api/transaction

Note: since forwarding such requests requires an additional cluster-internal HTTP request, they should be avoided when possible for best performance. Typically this is accomplished either by directing the requests to the correct Coordinator at a client-level or by enabling request “stickiness” on a load balancer. Since these approaches are not always possible in a given environment, we support the request forwarding as a fall-back solution.

Note: some endpoints which return “global” data, such as GET /_api/tasks will only return data corresponding to the server on which the request is executed. These endpoints will generally not work well with load-balancers.

Overload control

Introduced in: v3.9.0

arangod returns an x-arango-queue-time-seconds HTTP header with all responses. This header contains the most recent request queueing/dequeuing time (in seconds) as tracked by the server’s scheduler. This value can be used by client applications and drivers to detect server overload and react on it.

The arangod startup option --http.return-queue-time-header can be set to false to suppress these headers in responses sent by arangod.

In a cluster, the value returned in the x-arango-queue-time-seconds header is the most recent queueing/dequeuing request time of the Coordinator the request was sent to, except if the request is forwarded by the Coordinator to another Coordinator. In that case, the value will indicate the current queueing/dequeuing time of the forwarded-to Coordinator.

In addition, client applications and drivers can optionally augment the requests they send to arangod with the header x-arango-queue-time-seconds. If set, the value of the header should contain the maximum server-side queuing time (in seconds) that the client application is willing to accept. If the header is set in an incoming request, arangod will compare the current dequeuing time from its scheduler with the maximum queue time value contained in the request header. If the current queueing time exceeds the value set in the header, arangod will reject the request and return HTTP 412 (precondition failed) with the error code 21004 (queue time violated). Using a value of 0 or a non-numeric value in the header will lead to the header value being ignored by arangod.

There is also a metric arangodb_scheduler_queue_time_violations_total that is increased whenever a request is dropped because of the requested queue time not being satisfiable. Administrators can use this metric to monitor overload situations. Although all instance types will expose this metric, it will likely always be 0 on DB-Servers and Agency instances because the x-arango-queue-time-seconds header is not used in cluster-internal requests.

In a cluster, the x-arango-queue-time-seconds request header will be checked on the receiving Coordinator, before any request forwarding. If the request is forwarded by the Coordinator to a different Coordinator, the receiving Coordinator will also check the header on its own. Apart from that, the header will not be included in cluster-internal requests executed by the Coordinator, e.g. when the Coordinator issues sub-requests to DB-Servers or Agency instances.

Respond to liveliness probes

Introduced in: v3.10.0

By default, the HTTP REST interface of an arangod instance is opened late during the startup sequence. The instance responds with HTTP 503 (Service unavailable) until all REST APIs are available and usable.

You can optionally start the HTTP REST interface early in the startup sequence by setting the --server.early-connections startup option to true. This configuration allows an instance to respond to a limited set of REST APIs during the startup, even during the recovery procedure. This can be useful because the recovery procedure can take time proportional to the amount of data to be recovered.

The following APIs can reply early with an HTTP 200 status:

  • GET /_api/version and GET /_admin/version: These APIs return the server version number, but can also be used as a liveliness probe, to check if the instance is responding to incoming HTTP requests.
  • GET /_admin/status: This API returns information about the instance’s status, including the recovery progress and information about which server feature is currently starting.

During the early startup phase, all APIs other than the ones listed above are responded to with an HTTP response code 503, so that callers can see that the instance is not fully ready yet.

If --server.authentication is enabled, then only JWT authentication can be used during the early startup phase. Incoming requests relying on other authentication mechanisms that require access to the database data (e.g. HTTP basic authentication) are also responded to with HTTP 503 errors, even if correct credentials are used. This is because access to the database data is not possible early during the startup.

The GET /_admin/status API now also returns startup and recovery information. This can be used to determine the instance’s progress during startup. The new progress attribute will be returned inside the serverInfo object with the following sub-attributes:

  • phase: Name of the lifecycle phase the instance is currently in. Normally one of "in prepare", "in start", "in wait", "in shutdown", "in stop", or "in unprepare".
  • feature: Internal name of the feature that is currently being prepared, started, stopped or unprepared.
  • recoveryTick: Current recovery sequence number value, if the instance is currently recovering. If the instance is already past the recovery, this attribute contains the last handled recovery sequence number.

The exact values of these attributes should not be relied on, i.e. client applications should not check for any exact values in them. Feature and phase names are subject to change between different versions of ArangoDB. The progress attributes can still be used to determine whether the instance has made progress between two calls: if phase, feature, and recoveryTick don’t change, then there hasn’t been progress. Note that this is only true if the instance is still starting up. Once the instance has fully started and has opened the complete REST interface, the values in the progress attribute are expected to not change until shutdown.

Note that the maintenance attribute in responses to GET /_admin/status can still be used to determine whether the instance is fully available for arbitrary requests.