Unified Map to Define Job Interface

Unified Map to Define Job Interface

https://blueprints.launchpad.net/sahara/+spec/unified-job-interface-map

This specification proposes the addition of an “interface” map to the API for job creation, such that the operator registering a job can define a unified, human-readable way to pass in all arguments, parameters, and configurations that the execution of that job may require or accept. This will allow platform-agnostic wizarding at the job execution phase and allows users to document use of their own jobs once in a persistent, standardized format.

Problem description

At present, each of our data processing engines require or may optionally take any of arguments, parameters, configuration values, and data sources (which may be any of 0, 1, or many inputs to 0, 1, or many outputs). This forces users to bear the burden of documenting their own jobs outside of Sahara, potentially for job operators who may not be particularly technical.

A single, human readable way to define the interface of a job, that can be registered at the time of job registration (rather than job execution,) would allow several benefits:

  • A more unified UI flow across plugins
  • A clean separation of responsibility between the creator of the job (likely a technical user) and the executor of the job
  • A means of correcting our current assumptions regarding data sources (where for several plugins we are inappropriately assuming 1 input and 1 output source)

Proposed change

When creating a job, an optional “interface” list may be added to the job json (though we are fundamentally creating a map, presenting a list structure will allow more intuitive ordering and fewer unnecessary error cases.) Each member of this list describes an argument to the job (whether it is passed as a configuration value, a named argument, or a positional argument.)

The interface is described by the following jsonschema object field:

"interface": {
    "type": "array",
    "uniqueItems": True,
    "items": {
        "type": "object",
        "properties": {
            "name": {
                "type": "string",
                "minLength": 1
            },
            "description": {
                "type": "string"
            },
            "mapping": {
                "type": "object",
                "properties": {
                    "type": {
                        "type": "string",
                        "enum": ["args", "configs", "params"]
                    },
                    "location": {
                        "type": "string",
                        "minLength": 1
                    }
                },
                "additionalProperties": False,
                "required": [
                    "type",
                    "location"
                ]
            },
            "value_type": {
                "type": "string",
                "enum": ["string",
                         "number",
                         "data_source",
                         "input_data_source",
                         "output_data_source"],
                "default": "string"
            },
            "required": {
                "type": "boolean"
            },
            "default": {
                "type": "string"
            }
        },
        "additionalProperties": False,
        "required": [
            "name",
            "mapping",
            "required"
        ]
    }
}

Post-schema validations include:

  1. Names must be unique.
  2. Mapping is unique.
  3. The set of all positional arguments’ locations must be an unbroken integer sequence with an inclusive minimum of 0.
  4. Positional arguments may not be required, but must be given default values if they are not.

The job execution will also have a simpler interface field definition, described by:

"interface": {
    "type": "simple_config"
}

New error cases at execution time include:

  1. One configuration value or parameter is given two definitions (one through the interface map and one via configs, params, or data sources.)
  2. An interface value does not pass validation for the type specified for the field in question.
  3. A key in the execution interface map does not equal any key in the job definition’s interface map.
  4. The specified mapping type is not accepted by the job type being created (for instance, specifying the params type for a Spark job.)
  5. An input data source does not contain data.
  6. An output data source contains data.

In the case of additional positional values, the positional arguments given in the args list will be appended to the list of interface positional arguments (whether provided or default values.) This will allow for an *args pattern, should a plugin permit it.

Params and configs passed via the current mechanism that do not overlap with any key in the execution interface map will be merged and passed to the job as normal. This also applies to $INPUT and $OUTPUT params passed via the input source and output source fields.

Alternatives

In truth, after discussion, it seems that there is not a good alternative to the broad strokes of this plan (save doing nothing). Leaving all configuration of jobs to the execution phase is a real difficulty given that our supported data processing engines simply lack a unified interface. If we wish to create a unified flow, we need to create one; if we want to create one, the job definition phase produces the least user pain, and a simple, flat map is the most legible and flexible thing that can do the job.

Data model impact

A new table will need to be created for storage of interface fields, described by the following DDL (rendered in MySQL syntax for friendliness):

CREATE TABLE job_interface_arguments (
    id VARCHAR(36) NOT NULL,
    job_id VARCHAR(36) NOT NULL,
    name VARCHAR(80) NOT NULL, # ex: 'Main Class'
    description TEXT, # ex: 'The main Java class for this job.'
    mapping_type VARCHAR(80) NOT NULL, # ex: 'configs'
    location TEXT NOT NULL, # ex: 'edp.java.main_class'
    value_type VARCHAR(80) NOT NULL, # ex: 'string'
    required BOOL NOT NULL, # ex: 0
    order TINYINT NOT NULL, # ex: 1
    default_value TEXT, # ex: 'org.openstack.sahara.examples.WordCount'
    created_at DATETIME,
    PRIMARY KEY (id),
    FOREIGN KEY (job_id)
        REFERENCES jobs(id)
        ON DELETE CASCADE
);

This table will have uniqueness constraints on (job_id, name) and (job_id, mapping_type, location).

Note: While the TEXT type fields above (save Description) could validly be given an upper length limit and stored as VARCHARs, TEXT is safer in the case that a job actually requires an overly long argument, or is configured with a reasonably massive key. This implementation detail is certainly up for debate re: efficiency vs. usability.

Happily, this change will not require a migration for extant data; the interface fields table has a (0, 1, or many)-to-one relationship to the jobs table, and the existing configs/params/args method of propagating job execution data can continue to function.

REST API impact

The Create Job schema will have a new “interface” field, described above. Each listed exceptional case above will generate a 400: Bad Request error.

This field will also be represented in all GET methods of the Job resource.

The Create Job Execution schema will have a new “interface” field, described above. Each listed exceptional case above will generate a 400: Bad Request error. This field will not be returned on a GET of a job execution object; instead, the final, merged configuration will be returned.

No other impact is foreseen.

Note: I am profoundly open to better options for terminology throughout this document. As “args”, “params”, and “configs” are already taken, naming of a new option has become difficult. “Interface” and “Interface arguments” seem to me to be the best option remaining in all cases. If you can do one better, please do.

Note: As interface fields will be represented in the data layer as individual records, it would be possible to create an entirely new set of CRUD methods for this object. I believe that course of action to be unnecessarily heavy, however: should the job binary change, the job must be recreated regardless, and a sensible interface need not change for the life of any concrete binary.

Other end user impact

python-saharaclient will require changes precisely parallel to the interface changes described above.

Deployer impact

None.

Developer impact

None.

Sahara-image-elements impact

None.

Sahara-dashboard / Horizon impact

The immediate change does not require a Horizon change. Any UI that utilizes this feature should be represented as a separate blueprint and spec, and will doubtless touch wizarding decisions which are wholly orthogonal to this feature.

Implementation

Assignee(s)

Primary assignee:
egafford
Other contributors:
None

Work Items

  1. API updates as specified.
  2. DB layer updates as specified.
  3. Job execution argument validation and propagation to clusters.
  4. Testing.
  5. Python-saharaclient updates and testing.

Dependencies

None at present.

Testing

A tempest test will cover injection of each mapping type into jobs (args, configs, params.) This will be tested via a Pig job, as that type may take all of the above. This test will include arguments mapping to both a Swift datasource and an HDFS datasource, to ensure that both URL types are preserved through the flow.

Thorough unit testing is assumed.

Documentation Impact

None that have not already been mentioned.

References

Chat (2014/12/05; begins at 2014-12-05T16:07:55)

Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.

sahara-specs 0.0.1.dev331