Compute Jobs Reference

This is the reference documentation for Manta compute jobs. Unless otherwise specified, the semantics described here are stable, which means that you can expect that future updates will not change the documented behavior. You should avoid relying on behavior not specified here.

You should be familiar with the Manta storage service before reading this document. To learn the basics, see Getting Started.

Jobs Overview


Before getting started, it's worth describing some key terms:

Guiding principles

Several principles guide the design of the service:

On a per-operation basis, jobs provide the same strong consistency that the storage system does, but since jobs are composed of many individual operations, it's possible to see inconsistent results if the underlying data changes while the job is running. See "Processing inputs" for details.

In addition, jobs gain additional availability by retrying operations in the case of internal failures. So while a PUT may experience a 500 error during certain failure conditions, job tasks will generally be retried in the face of such failures, resulting in higher success rate at the cost of increased latency. See "Failures and internal retries" for details.

System Scale

In order to scale to very large data sets, there are no _hardcoded_ limits on:

In practice, object sizes are limited by the maximum size file that can be stored on storage servers as well as the largest practical HTTP request size.

The other quantities mentioned above are intended to be limited primarily to the physical resources available to Manta rather than architectural limits. Task stdout and stderr are limited to the amount of local disk space requested for each task. (See "Advanced output" for details and ways to emit much larger outputs.) The system never operates on the full set of tasks, inputs, outputs, or errors at once, nor does it provide APIs to do so until after the job has finished running. In practice, input and task counts over one hundred thousand in a single job currently can result in poor performance for other jobs.

Some dimensions of the system are explicitly limited (e.g., the total number of reducers, and the maximum number of phases in a job), but the established limits are intended to be sufficient for reasonable usage. These limits are subject to change.

Job configuration

The job configuration is mostly in the "phases" property, which might look like this for a typical two-phase map-reduce job:

    "phases":  [ {
         * "exec" is the actual bash script to run for each task.
         * "type" is "map" by default, which means there's one task for each
         * input object.
        "exec": "grep -w main"
    }, {
         * Reduce tasks have a fixed number (instead of one per object), and
         * operate on a stream of all of their inputs concatenated together.
        "type": "reduce",
        "exec": "/var/tmp/stackvis/bin/stackvis",

        * Use two reducers (not commonly necessary -- here for
        * illustration).  See "Multiple reducers" below.
       "count": 2,

        * Ensure that we're using a modern image (not commonly necessary --
        * here for illustration).  See "Compute instance images" below.
       "image": ">=13.1.0",

         * This job uses an asset that bundles the actual program to execute
         * along with its supporting data files.  We use an "init" script to
         * unpack the tarball before any tasks run.  See "Assets" below.
        "assets": [ "/:login/stor/stackvis.tgz" ],
        "init": "cd /var/tmp && tar xzf /assets/:login/stor/stackvis.tgz",

         * Request additional memory and disk resources (at additional
         * cost).  See "Resource usage" below.
        "memory": 2048,      /* megabytes */
        "disk": 16           /* gigabytes */
    } ]

This is a complex example to illustrate the available properties. Most jobs will only use a few properties, and many jobs will only use "exec".

Phase property reference

NameSummaryDefaultSee section
type"map" or "reduce""map""Tasks and task types"
execscript to run per tasknone"Task lifetime, success, and failure"
assetsextra objects to make available[]"Assets"
initscript to run before tasksno script"Initializing compute instance"
memoryrequested DRAM cap (MB)1024"Resource usage"
diskrequested disk quota (GB)8"Resource usage"
countnumber of reducers1"Multiple reducers"
imageimage version dependency"*""Compute instance images"

Tasks and task types ("type" property)

The actual work of jobs is divided into tasks, each representing one invocation of the user script. Tasks are executed inside operating system containers called "compute instances" that are isolated from each other and the rest of the service.

For map phases, there's one task per input object. Each map task has read-only access to a local file representing the contents of that input object, and stdin is redirected from that object. Map tasks run as soon as resources are available and are not blocked on the rest of the system once they start running.

For reduce phases, the number of tasks is configured statically by the "count" property, which defaults to 1. Stdin for reduce tasks is a pipe to which the contents of all of the reducer's input objects will automatically be concatenated and written. Reduce tasks generally don't start running until either the first input becomes available or the end-of-input is reached, and they may block once they start until more input becomes available or end-of-input is reached. Reduce tasks run even when given zero inputs. Reduce tasks have no direct access to inputs as files, since the objects may not be stored locally, will not all be ready when the reduce task starts running, and are not assumed to fit on a single server.

You'll notice that unlike other map-reduce services, Joyent Manta Storage Service tasks operate at the granularity of a complete object. If you want to process key-value pairs or something more structured, your script is responsible for parsing the input. Many Unix tools natively parse whitespace-delimited text formats, many other tools are available to parse other formats (e.g., json), and you can always use custom programs (see "Assets" below).

Task lifetime, success, and failure ("exec" property)

A task begins when the "exec" script is invoked with "bash -c". Stdin is redirected either from a local file (map) or a pipe (reduce) as described above. Stdout and stderr are always redirected to local files for capture by the system.

The task normally ends when that shell process exits, and the result is successful if that process exits with status 0. If the task succeeds, any output is forwarded on to subsequent job phases (if any) or becomes final job output (if not). See "Advanced output" below for details.

If the task fails (either because the process exits with non-zero status or because it or any of its child processes dump core), output is saved but not forwarded on. Stderr and any core file generated are also saved. A UserTaskError is emitted with a reference to the saved stderr and core file.

When the user process exits, any child processes it created are killed immediately. It's recommended that the first child process not exit until it has first waited for any child processes that it has created to exit.

If any of these processes dumps core, the task ends as a failure, regardless of what other processes are still running and regardless of the exit status of the first process.

Where possible, the system may run multiple tasks from the same job and phase sequentially in the same compute instance, so user scripts must be able to handle state left over from previous tasks from the same phase.

Assets ("assets" property)

Many tasks require programs, configuration files, or data files, which can be stored as objects but are not part of the actual data stream. You can download these explicitly in an "init" script, but for convenience you can also specify objects in a per-phase "assets" array. Assets will be downloaded into the compute instance before any "init" script is run (and so before any "exec" script is run as well).

Initializing compute instances ("init" property)

It's sometimes useful to perform some expensive setup operation once per compute instance so that the tasks that run in that instance can assume that operation has already been done. A good example is unpacking a tarball that contains the program files to be executed during the task. You can do this with an "init" script, which is executed with "bash -c" exactly once in each compute instance before any tasks are run.

A common pattern is to bundle a Node.js or Python program as a tarball, specify the tarball as an asset, use an "init" script to unpack the tarball once for each compute instance, and run an executable from the unpacked tarball in the "exec" script, as in:

"phases": [ {
    "assets": [ "/:login/stor/stackvis.tgz" ],
    "init": "cd /var/tmp && tar xzf /assets/:login/stor/stackvis.tgz",
    "exec": "/var/tmp/stackvis/stackvis"
} ]

If this script fails for any of the reasons that tasks fail (see below), the tasks that would be run in that instance may fail with a TaskInitError.

The resources used by "init" scripts are charged to an arbitrary task that will be run in that compute instance.

Resource usage ("memory" and "disk" properties)

There are no explicit limits on CPU usage, network utilization, or disk I/O utilization, but use of those resources may be limited based on availability. This may vary across different tasks in the same job or tasks in different jobs.

By default, compute instances are given caps of 1024MB for both resident set (memory) and anonymous memory. The "memory" property of a phase allows users to request more of both, as either 2GB, 4GB, or 8GB. If that memory is available, the task gets it. Otherwise, the task may be queued until the memory becomes available or it may fail with a TaskInitError if the service determines that it's unlikely to have memory available any time soon.

Similarly, by default compute instances are given caps of 8GB of disk space. The "disk" property of a phase allows users to request more space in GB, as either 16GB, 32GB, 64GB, 128GB, 256GB, 512GB, or 1TB. As with memory, if the space is available, the task gets it. Otherwise, the task may be queued until the disk becomes available or it may fail with a TaskInitError if the service determines that it's unlikely to have disk available any time soon.

Note: each task's stdout and stderr are staged to the local disk for the duration of the task. For programs that emit a lot of data to stdout or stderr, the "disk" property may need to be adjusted accordingly.

These defaults are not stable. If you intend to depend on these values, you should explicitly specify "1024" for memory and "8" for disk.

Multiple reducers ("count" property)

Sometimes it's necessary to pipeline the reduce phase so that instead of processing all input objects in one pass, the input objects are processed in chunks across multiple passes. For example, you may want to process half of the inputs in each of two parallel tasks, and then process the output of that, to avoid having to load an entire data set in memory at once.

To support multiple reducers in parallel (i.e., in the same job phase), use the "count" property:

"phases": [
    ... /* any number of map phases */
    , {
        "type": "reduce",
        "exec": ...,
        "count": 2
    }, {
        "type": "reduce",
        "exec": ...

In this example, some number of map phases are followed by a phase with two reducers running in parallel, followed by a final reduce stage. The two parallel reducers are identical except for their input and output objects.

By default, inputs are randomly assigned to reducers. If a more sophisticated assignment is necessary (e.g., to ensure that certain groups of inputs are processed by the same reducer), the previous map phase can specify to which reducer a given output object will be assigned using the "-r" flag to mpipe, which must be an integer between 0 and N - 1, where N is the total number of reducers in the next phase. Also see documentation for "msplit".

Compute instance images ("image" property)

Compute instances are Joyent instances based on the manta-compute image. This image is essentially a base image with nearly all of the available packages preinstalled.

By default, tasks run in compute instances with the most recently released image available. You can specify a particular range of image versions using a semver-like value for the "image" property (e.g., "13.1.*"). If that image is not available, tasks for that job will fail with an InvalidArgumentError.

You're strongly discouraged from depending on an exact version of the image, as Joyent may frequently release minor updates to existing images and retire older versions as long as the new one is backwards-compatible. Dependencies are intended so you can tie jobs to major releases, or "at least" a particular minor release.

If you want to test your scripts and binaries in a Manta environment, use the mlogin command from your workstation. See Try your job with mlogin later in this document.

Task environment and authentication

For convenience, several environment variables are set for the "exec" script:

You can use the normal CLI tools (e.g., mls, mput, mget), which will use the private MANTA_URL endpoint. Requests to this endpoint are implicitly authenticated as the user under whom the job is running, though no private keys are available inside the compute instance. You can also use other tools that use the MANTA_URL and MANTA_USER environment variables.

Of course, if you want to make requests as another user for which you have the private key available, you can do so by overriding the appropriate environment variables, including MANTA_URL.

Advanced output

The task's stdout and stderr are redirected to a local file on disk. By default, if the task succeeds, this file is uploaded and becomes the task's sole output, which is forwarded to the next phase in the job (if any) or else become outputs of the job itself (if not). Intermediate objects are not directly exposed to you, and may never even be stored, but final outputs are always objects. Because the stdout and stderr are staged to disk, these are limited in size by the amount of local disk space available, which is controlled by the "disk" property of the task's phase. See the documentation on the "disk" property for details.

Several tools are available in compute instances for more sophisticated types of output. In addition, users can make use of advanced input using the HTTP API directly, which allows for emitting outputs not limited by the amount of local disk space available. See "Using the HTTP API for advanced output" below.

mpipe: advanced output


mpipe [-p] [-r rIdx] [-H header:value ...] [manta path]

Each invocation of mpipe reads data from stdin, potentially buffers it to local disk, and saves it as task output. If a Joyent Manta Storage Service path is given, the output is saved to that path. Otherwise, the object is stored with a unique name in the job's directory. If -p is given, required parent directories are automatically created (like "mkdir -p").

If you use mpipe in a task, the task's stdout will not be captured and saved as it is by default.

As a simple example,

wc | mpipe

is exactly equivalent to just:


since both capture the stdout of wc and emit it as a single output object. But you use mpipe for several reasons:

mcat: emit objects by reference


mcat FILE ...

mcat emits the contents of a object as an output of the current task, but without actually fetching the data. For example:

mcat ~~/stor/scores.csv

emits the object ~~/stor/scores.csv as an input to the next phase (or as a final job output), but without actually downloading it as part of the current phase.

As with mpipe, when you use mcat, the task's stdout will not be captured and saved as it is by default.

mcat is particularly useful when you tend to run many jobs on the same large set of input objects. You can store the set of objects in a separate "manifest" object and have the first phase of your job process that with "mcat". So instead of this:

mfind ~~/public | mjob create -m wc

which may take a long time if mfind returns a lot of objects, you could do this once:

mfind ~~/public > /var/tmp/inputs
$ mput -f /var/tmp/inputs ~~/public/inputs

And then for subsequent jobs, just do this:

echo ~~/public/inputs | mjob create -m "xargs mcat" -m wc

This is much quicker to kick off, since you're just uploading one object name. The first phase invokes "mcat" on lines from ~~/public/inputs. Each of these lines is treated as a path, and the corresponding object becomes an input to the second phase.

The object path is not resolved until it's processed for the next phase. So if you specify an object that does not exist, this will produce a ResourceNotFoundError for the phase after the mcat. Similarly, if you specify an object that you don't have access to, you'll get an error in the next phase when you try to use it.

msplit: demux a stream for reducers

Reads content from stdin and outputs to the number of mpipe processes for the number of reducers that are specified. The field list is an optional list of fields that are used as input to the partitioning function. The field list defaults to 1. The delimiter is used to split the line to extract the key fields. The delimiter defaults to (tab). For example, this will split stdin by comma and use the 5th and 3rd fields for the partioning key, going to 4 reducers:

msplit -d ',' -f 5,3 -n 4

mtee: save stdout to an object in a stream of commands

mtee is like mput, but takes input on stdin instead of a file, and emits its input on stdout as well, much like tee(1).

mtee is also similar to mpipe, except that the newly created object does not become an output object for the current task, and using mtee does not prevent stdout from being captured.

For example, this will capture the output of cmd to an object ~~/stor/tee.out and still pipe what was coming from cmd to cmd2:

cmd | mtee ~~/stor/tee.out | cmd2

Using the HTTP API for advanced output

The functionality provided by "mpipe" and "mcat" can be accessed directly using any HTTP client. Clients can emit output objects using normal PUT operations using the parameters specified by the MANTA_URL, MANTA_USER, and MANTA_NO_AUTH environment variables. (See "Task environment and authentication" for information about using these variables inside a job.) Since the other tools buffer object contents to local disk, this approach is necessary to emit objects whose size is not bounded by the local disk space available. These requests are subject to the same timeouts as normal requests to the public Manta endpoints, and they may be terminated abruptly if the data stream is idle for more than a minute.

Job-specific behavior is controlled by several headers:

Since these headers are specified in a PUT, you can name your output object whatever you want. The MANTA_OUTPUT_BASE environment variable is provided as a unique, relatively friendly base name for output objects. See "Task environment and authentication for details.

These headers are only interpreted by the server located at $MANTA_URL inside the context of a job. You cannot use these headers outside of a job, nor can you use them with the public API servers.

For example,

... | mpipe

is analogous to PUT $MANTA_URL/$MANTA_OUTPUT_BASE.$(uuid) with "X-Manta-Stream: stdout".


... | mpipe -r2 /:login/stor/output.txt

is analogous to PUT $MANTA_URL/$MANTA_USER/stor/output.txt with "X-Manta-Stream: stdout" and "X-Manta-Reducer: 2".


mcat /:login/stor/input.txt

is analogous to PUT $MANTA_URL/$MANTA_USER/stor/input.txt with a content-length of 0, no content, and "X-Manta-Reference: true".

Job execution

When a job is first submitted, its state is "queued". Under normal conditions, the job immediately transitions to the "running" state, but since no inputs have been submitted, there are no map or reduce tasks running.

Processing inputs

As job inputs are submitted (typically with mjob create or mjob addinputs), the objects are located within the service and dispatched for the first phase. For a map phase, dispatching means issuing a new map task to be executed on one of the physical servers that stores the object. For a reduce phase, dispatching means feeding the input to the stdin of one of the reducers.

If the input object cannot be found, a ResourceNotFoundError is emitted. Access control is enforced for jobs just like for GETs. If a GET for an object would return an error because the user doesn't have access to the object, the same error will be emitted if the user attempts to process that object as part of the job.

If an object is removed while there's a job operating on it, that job may successfully process the object or it may issue a ResourceNotFoundError. Either behavior is possible regardless of whether the deletion completes before, during, or after the task that would process that object starts or finishes.

Executing tasks

As tasks are issued, they may begin executing or they may queue up, depending on service load. As they execute, they emit one or more output objects.

When any task fails, its outputs are saved along with its stderr and up to 1 core file. A UserTaskError is emitted that references the stderr and core file.

When a final-phase task completes successfully, its outputs are saved and marked as job outputs.

The outputs of non-final-phase tasks are called intermediate objects. When such tasks completes successfully, these intermediate objects are dispatched to the next phase similar to the way job inputs are dispatched: if the next phase is a map phase, a map task is issued for each intermediate object (which may immediately begin executing), and otherwise the intermediate object is fed as an input to one of the reducers of the next phase (which may immediately begin processing it). If necessary, you can specify which reducer the output should be sent to using mpipe. See "Advanced output" for details.

Notably, this means that job phases are not serialized. You can stream input to a three-phase map job and have tasks running for all three phases. You can have final outputs available for some of the first inputs before you've finished submitting all of the inputs.

Ending input

When the user finally ends the job input stream (which happens automatically with mjob create or mjob addinputs unless the "--open" flag is specified), any subsequently submitted input objects will not be processed.

The job's end-of-input will be propagated to the first reduce phase, if any. When each reducer finishes reading all of its input, it will read EOF. As all tasks in a given reduce phase complete, the end-of-input is propagated to the subsequent reduce phase. (End-of-input has no meaning for map phases.)

The input stream may automatically be ended for jobs that have no inputs added for an extended period (many minutes).

Job completion and archival

When all inputs of the job have been processed as described above, and all outputs from completed tasks have been propagated, and all tasks have been completed, then the job's state becomes "done".

Because jobs can generate an enormous amount of data, and the system cannot predict how long you want to keep this data, the lists of inputs, outputs, and errors of your job are automatically converted to flat objects after the job reaches the "done" state. Once this archival process completes, interacting with your job starts to look exactly like interacting with directories and objects: your job is a directory, and archived data are just regular objects. It is your responsibility to delete all archived job data when you're through with it; until then, you will be billed for it.

Once a job is archived, the /jobs/:id/live/* APIs will continue to function for at least one hour, but at some point after that, they will start returning 404, after which only the archived objects will be available.


Jobs may be cancelled any time while they are still running. No new tasks will be dispatched, and running tasks will be cancelled as soon as possible, which may result in errors being reported. The job's state will become "done", though there may still be some tasks running at this point.

Under some conditions, job cancellation may result in intermediate objects being left around under the job's directory. Users should perform a recursive remove ("mrm -r") of the job's directory to clean up these artifacts.

Cancellation is intended to be an exceptional case, and the stats, outputs, errors, and side effects of cancelled jobs are basically undefined.

Failures and internal retries

There are many internal operations executed as part of the job's execution, including object lookup, saving stdout for all tasks, and fetching inputs for reduce tasks. Where possible, these operations are automatically retried a small number of times, which should have no visible impact except for additional latency. (For example, retries saving stdout will not result in multiple "copies" of the output.) Errors will be emitted for operations that continue to fail.

Some internal failures may result in tasks being executed more than once. The service will ensure that only one of those attempts' outputs will be used, but you must still keep this in mind if your jobs produce side effects (e.g., write requests to external services). In short, any task may be executed multiple times, even successfully.

Failures of the user script are not ever retried. If the script returns non-zero, dumps core, or fails in some other way, the task fails immediately and is not retried. If your script sometimes fails transiently, and you want it to be retried, you must do this yourself at the appropriate level (which may be at the level of the whole script, or just part of it, depending on the application). Built-in tools like "mpipe" automatically retry a small number of times.

Debugging jobs

Debugging programs that run remotely poses challenges, but Manta provides several facilities to make this easier.

Save debug output to Manta

By default, Manta does not save the stderr emitted by your job, but if your program returns a non-zero exit status or dumps core, stderr is automatically saved and made available to you. You can use this to figure out what part of your program failed, and why.

You can also save your own output objects with mput. The mtee command is also available for inserting into a pipeline (like tee(1)). These approaches are difficult to scale up, but can work well for small jobs.

Try your job with mlogin

Since it's often easiest to debug a program by running it by hand from an interactive shell, the Manta CLI tools include an mlogin command, which starts an interactive shell inside a Manta job, directly where your object is stored. Here you can inspect the environment, run your program as many times as you want, save intermediate files, use a debugger, or do whatever else you need to debug your program.

Here's a simple example where the user (called "mantauser") interactively runs echo and wc:

mlogin /manta/public/examples/shakespeare/kinglear.txt
 * created interactive job -- 3226d090-9dde-4dc0-b59a-c80d59635c63
 * waiting for session... - established

mantauser@manta # echo $MANTA_INPUT_FILE

mantauser@manta # wc $MANTA_INPUT_FILE
  5525  27770 157283 /manta/manta/public/examples/shakespeare/kinglear.txt

Since mlogin runs in the exact same environment that normal Manta jobs run, once your program works reliably under mlogin, it will work as a non-interactive job as well.

Try your job locally

It's easy to forget that since Manta's environment is very similar to most other Unix-like systems (e.g., SmartOS, MacOS, and GNU/Linux), you can also test your program by hand on your own system. Once it works reliably there, it's likely it will run correctly in Manta as well.

Debugging reducers

You can use several of the above techniques for debugging the first phase of a job, but it's not obvious how to apply them to the second phase of map-reduce job (or even a map-map job), since the input is not available for you to use mlogin with. One solution is to replace your reduce phase with cat, which will simply copy the reducer's input as output. Then you can mlogin in to the job's output file and debug it. (This technique also works with multiple reducers in a single phase.)

For example, suppose you're having trouble with this two-phase map-reduce job, which runs wc in the first phase and a custom script in the second phase.

mjob create \
    -m wc \
    -s /path/to/myscript -r /assets/path/to/myscript

You would replace that job with:

mjob create \
    -m wc \
    -r cat

When that job completes, you'd run mlogin on the output of that job, which is exactly the input to your reducer. (If you had multiple reducers in this phase, there would be multiple outputs, one for each reducer, and you could debug each one separately.)

This process is analogous to debugging a Unix pipeline on your local machine by saving the intermediate output of the first part of the pipeline to a file first, and debugging the second part of the pipeline using that file.