dcp-rds

DCP Remote Data Service

This is the official reference implementation of a DCP Remote Data Service. While dcp-rds can be used as a production service, its purpose is primarily to demonstrate how to write a Remote Data Service service for DCP.

Record Of Issue

Date	Author	Change
Oct 8 2021	Wes Garland	Initial Release

Legal (MIT License)

Copyright (c) 2021 Kings Distributed Systems, Ltd.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Icon

The "Pizza Slice" icon associated with but not part of dcp-rds was made by turkkub from Flaticon.

About

DCP allows software developers who write DCP Client Applications to map an input set to an output set via a work function and a work argument. The typical data flow has the client sending the input set, work function, and work argument to the centralized scheduler, which mediates all communication with the workers. A DCP Remote Service provides a mechanism for clients and workers to communicate directly with data sources or sinks.

A DCP Remote Data Service consists of two parts,

1. a program which can receive data from a remote entity, and
2. a program which can send this data to clients and workers when they ask for it

For the reference implementation, we have chosen to write both #1 and #2 into a single Node.js program. We foresee many implementations where #1 and #2 are nothing more than ftpd and httpd with carefully-crafted file extensions and mime.types.

Interface

1. An HTTP POST is made to the server (dcp-rds-httpd), which records the content to disk. The server responds with HTTP status 200, content-type application/json; the response body is a JSON-stringified object.
2. Workers make an HTTP GET request to dcp-rds-httpd, which responds with the content on disk.
3. Content stored within the dcp-rds is uniquely identified by the tuple (job, elementType, element).

methods/upload

POST Parameters

job

This is a unique identifier for the job. It can be any unique string that describes the data that the job is for. Future versions of DCP will support UUIDs which are calculated for jobs before deployment; this would be a good choice for this parameter.

elementType

The type of element that is stored within the job, for example 'slice', 'argument', 'work-function', etc.

element

Combined with elementType, this is a unique identifier for the content within the job. For example, if the elementType is 'slice', a good element might be the sliceNumber.

content

This is the content to store that corresponds to the tuple (job, elementType, element). The uploaded datum will not be interpreted in any way; the worker will receive exactly the same bytes for this slice as the dcp-rds-httpd daemon did. The content-type will also be preserved from the upload, however, note that is it possible for both this server and intermediary web servers and caches to transparently alter the content transfer encoding, etc.

contentType

The MIME type of the content, after it has been decoded from the form upload.

Response

Successful

Property	Value
success	true
href	a full URL which gives the path from which this datum may be downloaded

Failure

Property	Value
success	false
error	optional instance of Error. If present, may have also have non-standard stack and code properties.

Integration with compute::RemoteDataPattern

When using compute::RemoteDataPattern to store input data, we recommend using the elementType 'slice' and using the sliceNumber for the element during the upload. This will all the datum to be in a deterministic (pattern) location, given by ${website origin}/methods/download/jobs/${job}/slice/${slice}

Note: As of this writing (Oct 2021), developers currently need to generate their own unique jobIds and use a per-job pattern with it hard-coded within the pattern, as there is no way to get a job Id from the scheduler before deployment. This will be addressed in a future version of the Compute API.

Integration with compute::RemoteDataSet

In all cases, the response indicates the location from which the uploaded datum can be retrieved.

When either the job or element parameters are missing, the remote data service will generate a random identifier to take its place. This allows the uploading of content whichhas not yet been associated with a job, but makes it imperative that the href URL in the reponse is memoized so that the content can be located later when the job is deployed.

methods/upload-set

Note - As of this writing (Oct 2021), this method has not been implemented.

POST Parameters

job

See: methods/upload

content

Specifies the entire input set for this job and must be an Object which has been serialized either with JSON or KVIN. Serialization type will be automatically detected, but any data type which requires application/x-kvin output (eg TypeArrays, see table below) must be uploaded in KVIN to preserve type information correctly.

The service will intelligently decide how to store the data on disk, and will serialize it for transmission to the worker using whatever serialization is appropriate for the content type.

contentType

The content type of the data will be automatically determined after it has been extracted from the uploaded Array, evaluated on a per-element basis. If the initial serialization is done correctly, the JS type used in the uploaded Array will be the same as the JS type the Worker receives as its work function argument, or slice datum.

The following table lists the internal representation which is used to achieve this, but should be considered an implementation detail. The table is traversed in top-to-bottom order, with the first match being used.

JS Type	MIME Type
string	text/plain
number ∈ (-Infinity, Infinity)	application/x-json
number	application/x-kvin
Uint8Array	application/octet-stream
TypedArray	application/x-kvin
other	application/x-kvin

Response

Successful

Property	Value
success	true
pattern	a pattern string which can be used to retrieve this data, suitable for use with compute::RemoteDataSet
length	number of elements in the request

Failure

Property	Value
success	false
error	optional instance of Error. If present, may have also have non-standard stack and code properties.

Content-Types

DCP workers uses the HTTP content-type header to sort out what kind of serialization (if any) is in use for a given HTTP request. (See: fetchURI())

Important: the contentType of the data in the POST request is generally not the same as the Content-Type header of the POST request; if you are developing a content uploader, it is important not to conflate the two!

MIME Type	Return Type
text/plain or none	string primitive
text/*	An instance of String with the contentType property set, and the charset property set if it was specified
application/json	whatever JSON.parse returns on the data when it is treated as a string
application/x-kvin	whatever kvin.deserialze returns on the data when it is treated as a string
image/*	A Uint8Array of the decoded contents with the contentType property set.
application/octet-stream	A Uint8Array of the decoded contents with the contentType property set and the charset property set if it was specified.
Anything else	Same as application/octet-stream

Installing

dcp-rds-httpd is a simple web server that can communicate directly over HTTP, but the typical way system administrators deploy this product is to have it listen to localhost and communicate with the outside world via proxy with a full-featured web server, such as nginx or Apache httpd. These servers allow administrators to configure access control, TLS certificates, sophisticated logging, etc., as well as interoperation with local firewalls and other content on the same host.

Do not expose dcp-rds-httpd directly to the internet without access controls in place!!!

Prerequisites

• Ubuntu Linux 20.04
• Node.js 12+
• nginx or Apache httpd
• Knowledge of organizational security policies and ability to implement them with Apache or nginx.

Common

PREFIX="/var/dcp-rds"

Package install

sudo useradd -d "${PREFIX}" -M -s /bin/false dcp-rds
sudo mkdir -p "${PREFIX}" 
sudo chown dcp-rds:dcp-rds "${PREFIX}"
sudo npm i dcp-rds --global --prefix="${PREFIX}"

Systemctl Integration

sed -e "s;/var/dcp-rds/;${PREFIX}/;" \
< "${PREFIX}/systemctl/dcp-rds.service" \
| sudo sh -c 'cat > /etc/systemd/system/dcp-rds.service'
sudo systemctl daemon-reload
sudo systemctl restart dcp-rds

nginx Integration

Apache Integration

• now edit apache confs
• copy systemctl stuff
• systemctl daemon-reload
• systemctl start whatever this thing is call

Appendix A - Manifest

Filename	Description
README.md	This document
bin/dcp-rds-httpd	Remote Data Service daemon; web server written in Node.js
etc/config.js	Configuration object
etc/local-config.js	If present, overrides properties of etc/config.js
libexec/	JS code in this directory can be executed on the local machine by a remote entity
browser-test/upload.html	Test harness - web page to upload data into the service
systemctl/	Sample configuration files for systemctl
apache/	Sample configuration files Apache httpd
nginx/	Sample configuration files for nginx

Appendix B - Remote URI Data Movement

Note: It is possible to freely mix remote URI and scheduler-centric URIs (traditional DCP) within the same job, at will, on an ad-hoc basis.

The dcp-rds-httpd program can be used for any of the server elements in this diagram:

sequenceDiagram 
participant client
participant scheduler
participant worker
participant sliceDataServer
participant workFunctionServer
participant argumentServer
participant resultStorageServer

client ->> scheduler: input data URIs
client ->> scheduler: work function URI
client ->> scheduler: work argument URIs

scheduler ->> worker: input data URIs
scheduler ->> worker: work function URI
scheduler ->> worker: work argument URIs

worker ->> sliceDataServer: datum URI
sliceDataServer ->> worker: datum

worker ->> workFunctionServer: work function URI
workFunctionServer ->> worker: work function

worker ->> argumentServer: argument URI
argumentServer ->> worker: argument

worker ->> scheduler: events (console.log)
scheduler ->> client: events (console.log)

worker ->> resultStorageServer: result
resultStorageServer ->> worker: resultURI
worker ->> scheduler: result URIs
scheduler ->> client: result URIs

Appendix C - Compute API support for Remote Data

• RemoteDataSet
• RemoteDataPattern
• JobHandle.setResultStorage

let job = compute.for(inputSet, new URL('http://my.site.com/work.js'));

let { RemoteDataPattern } = require('dcp/compute');
let myJobId = Math.random();
let pattern = 'http://dcp-rds.site.com/methods/download/jobs/${myJobId}/slice/{slice}';
let job = compute.for(RemoteDataPattern(pattern, 2), function(el) { console.log(el)});
job.on('console', console.log)
await job.exec();

job.setResultStorage(new URL('https://192.168.123.1/submit-result'), { juuid: job.uuid });