REGARDS processing microservice

Processing architecture

Domain entities

There are five main entities introduced by Processing. Since their names are rather generic and semantically overloaded, they are prefixed by a P referring to the "P"rocessing context.

Before describing each of the entities in more details, here is the very crude idea behind this architecture:

• a PProcess describes a program to be run on data, the program can be configured with parameters,
• a PBatch describes a logical group of executions of a PProcess, providing values for the program configuration parameters,
• a PExecution describes the launch of the program on data, providing the actual input files,
• a PInputFile the description of a file to be used during an execution (it is actually a glorified URL pointing to some data),
• a PStep is en event occuring during an execution (flagging the execution as running, in success or in failure),
• a POutputFile is a result of an execution, and may refer to the PInputFiles which were used to build it.

PProcess

At the top of the abstraction hierarchy is the PProcess. A PProcess is the abstract description of a procedure to be applied to one or many files.

The PProcess has several expected attribute:

• an UUID,
• a name,
• a set of metadata named the "process info".

It also provides more specific attributes:

• a size forecast, allowing to predict before launching execution, the expected quantity of data outputted for a given input size,
• a duration forecast, allowing to predict the expected duration of treatment, allowing to compute a timeout duration for each execution,
• a mapper from output to input, telling how to assign the corresponding inputs to a given output,
• validators for PBatch and PExecution creation using this PProcess, allowing to decide in advance if the process may be used in a given context,
• a workload engine and an executable, describing what and how to run the corresponding program ; these mechanism will be explained below.

PProcess is an interface, and there is a default basic concrete class ConcretePProcess implementing it as a simple POJO. However, any type of class could implement PProcess, depending on the final application.

PBatch

A user has data to process, and has chosen a PProcess to treat them. In order to launch treatments (also called executions), the user must first acquire a PBatch, by calling the Processing service and providing:

• the PProcess configuration parameter values to be used by all executions
• statistics regarding the quantity of data to treat.

This allows the Processing service to decide whether the treatments can be done or not. For instance, some configuration parameter values might be wrong, or some quota of use may be reached. This validation is done using the PProcess batch validator.

If the batch validator accepts the information, it delivers a PBatch instance, saved in the Processing service database. This batch has:

• an ID,
• a correlation ID given by the user when requesting the creation of the batch, which will be provided for any execution result later,
• the values for the PProcess parameters which will be used by all the executions launched in the context of this batch.

A batch is an immutable entity: it never changes once it has been created.

Abusing batches is possible

Once the batch has been created and delivered to the user, the user may use it and abuse it. Theere is for now no limitation of the number of executions a user can do using the same batch ID. This limitation is not enforced because sometimes, the input statistics provided at the batch creation may be slightly inexact, and it is hard to predict in advance how many executions will be needed to treat all the data. Instead of making bad heuristics, the choice has been made to trust the user to not abuse the system, especially since for now the only user for Processing is the Order service, which has no interest in abusing batches.

However, if Processing had to be accessed directly by end users in the future, it will require to enforce a more strict policy regarding the execution creation within a given batch.

PExecution

When the user has received the batch response providing the batch ID, the user can send execution request to Processing, using an AMQP PExecutionRequestEvent message.

The execution request contains:

• the batch ID it corresponds to (determining the configuration parameters for the PProcess),
• the list of PInputFile to use as input data for this execution,
• an execution correlation ID, which will be provided back when Processing returns the execution result (as an AMQP PExecutionResultEvent).

When the Processing service receives the request, it validates it using the PProcess execution validator, and if accepted, saves it in the service database and then runs it, using the PProcess executable and workload engine.

A Pexecution is a mutable entity: while running the executable, PSteps are emmitted and added to the PExecution, providing a current status and last update timestamp for the execution.

PInputFile

A PInputFile is an input file descriptor, defining:

• where the file is located using an URL,
• the file size/content type/checksum,
• a file-specific correlation ID, which will be used when reporting back the execution result, allowing to map output files back to the input file correlation IDs which were used to generate the output file.

A PInputFile is an immutable entity.

PStep

While running in the context of an execution, the process-supplied IExecutable has access to en event notifier, allowing to inform the Processing service that steps have occurred in the execution.

A PStep is an immutable entity with the following fields:

• a status defining the type of step (RUNNING, SUCCESS, FAILURE, etc.).
• an optional set of output files used if the step is final.

An execution status can be transitory of final. Final statuses are:

• SUCCESS,
• FAILURE,
• and TIMED_OUT.

Once a final status is reached, the execution is considered finished and the PExecutionResultEvent is created and sent to the AMQP broker.

POutputFile

When a PStep is final, it is supposed to contain a list of output results, in the form of POutputFiles.

A POutputFile is, like the PInputFile, a wrapper around an URL. It has:

• an ID,
• a reference to the corresponding execution UUID,
• a file name, URL, size and checksum,
• a list of optional PInputFile correlation IDs, allowing to make reference to the input which generated this output.

The POutputFile is mutable and also has a few technical flags:

• downloaded tells if the file has been downloaded by the client,
• deleted tells if the file has been deleted and is not present for download anymore.

General algorithm

Here is a description of the way a client needs to interact with Processing in order to launch executions. (Only the 'happy path' is represented.)

The client needs to interact first with the REST API in order to get the PProcess details and then create a PBatch.

Then, the user needs to send PExecution requests events through AMQP.

The Processing service then launches the work using the workload engine for the corresponding process, which launches an executable (provided by the process).

Running the executable generates steps, which are notified using a step notifier present in the execution context which is given to the executable.

When the notifier receives a final step (step with a final ExecutionStatus), it generates an execution result event sent back to the client though AMQP.

The final step contains references to the output files generated during the execution. The user is free to download them by whichever means it has. Once the download is done, the user notifies Processing through AMQP of the downloaded files. Processing marks them as downloaded.

A scheduled service runs at a given periodicity to take all the output files marked as downloaded, and delete them from the shared storage space where they have been stored, and marks the output files as deleted in the database.

A scheduled service runs at a given periodicity to take all the batches for which all the executions have been finished for a given duration, and deletes them from the database. This deletion cascades to all the executions, steps, and output files linked to this batch.

Technical details

Workload engine / executable / execution context

This section describes the why and how of the workload engine, and the mechanism by which executables notify execution events.

The Processing service is an empty shell. It does not run the processes itself, but delegates the running to the workload engine referenced by a PProcess.

Generalities on workload engines

The WorkloadEngine responsibility is then to define how the IExecutable defined by the process will be launched (which could be synchronously, in a separate thread, in a different virtual machine, etc.). A WorkloadEngine is a way to abstract the running of processes. It is an interface to frameworks like REGARDS jobs, Spring Batch, etc. whose purpose is to launch programs on demand.

A WorkloadEngine instance has a name, and references itself when loaded at bootstrap in a WorkloadEngineRepository. A PProcess will provide the name of the engine it is supposed to run on.

When an execution of the PProcess is started, the service will load the corresponding WorkloadEngine. For instance, in the case of the deployment of Processsing for launching treatments on orders, the WorkloadEngine used by all processes is for now based on REGARDS jobs. Launching an execution with this engine is just creating a LaunchingExecutionJob as QUEUED.

ExecutionEvent

An execution event signifies that something happened during an execution. The execution event contains an optional PStep (which may be final or not) and an optional list of POutputFiles. The event can signify that a step has been reached, or that output files have been generated, or both.

(In the context of the use of Processing for order data treatment, the processes must send output files only along with a final step, but this is a limitation given by the context of use, rather than a limitation of the domain as it is designed.)

IExecutionEventNotifier

The IExecutionEventNotifier interface defines how to deal with events sent during an execution. It is a simple functional interface which takes as input an ExecutionEvent and returns a Mono<PExecution> (with the steps potentially updated with the event step). (The Mono wrapper signifies that a side-effect has been done during the execution, for instance a modification of the execution in the database.)

ExecutionContext

An ExecutionContext provides access to:

• the PProcess and PBatch corresponding to the execution,
• an IExecutionEventNotifier,
• a mutable map of metadata.

IExecutable

An IExecutable is a function which takes as input an ExecutionContext and returns a Mono<ExecutionContext>. The Mono wrapper signifies that side effects may occur.

IExecutables are simple functions which can be chained, allowing for compositionality and modularity. Simple executable steps can be aggregated into more complex executables.

These functions are meant to do side effects, and use the input ExecutionContext's event notifier, in order to advance steps or signify that outputs have been generated.

An IExecutable is not supposed to know anything about the workload engine that runs it, but it practise some information can be passed by the workload engine to the executable through the ExecutionContext metadata.

Reactive-first service

The Processing microservice has been identified as a potential bottleneck with heavy load. Because it might also be often prompted for monitoring information, the choice has been made to make this microservice with a reactive stack, using Spring Webflux and Netty as the backbone.

However, in the context of REGARDS, and because there is a lot of security configuration inherited from REGARDS, the service can be (and has been) configured to run in a servlet context, backed by Jetty.

It should be beneficial to the service to remain in a reactive-first logic, especially in its domain core. The reactive model permeates throughout the code in the form of function from values to Monos and Fluxes of values. (Clean code often tends to leave such frameworks out of the domain, but it is very hard to do so with Reactor ; especially since the underlying interfaces do not discriminate between single-valued publishers and multi-valued ones, and typing everything as Publisher is not helpful.)

As for the REST controllers, two versions are provided: a reactive and a servlet one.

Clean architecture

The Processing service is based on 'Clean Architecture' principles. This means that it has a domain core and several adapters, which abstract away the different ways to interact with the external world.

The core defines:

• the base entities describes above,
• the interfaces for services and repositories to manage the entities.

The main adapters are:

• database access,
• event handlers,
• process management,
• workload engine.

Database access adapter

This adapter uses a reactive dirver (r2dbc) and is not destined to be reimplemented for another context, unless there is a need to connect to something else than postgresql.

In order to allow Processing to be run effectively in all contexts, and in order to allow imitations on PProcess, the choice has been made to use the reactive driver R2DBC to the database.

This allows to use Processing in a reactive context without suffering from thread-blocking in database calls.

This has several consequences however:

• reactive drivers to databases are bad at relations, which is why the database is de-normalized in several ways, attempts to re-normalize the schema will involve performance costs, and need for way more requests to the database,
• there is a clean border between the domain layer and the data access layer, meaning that there is a need for mapping domain entities to database entities.

In the database access layer, the objects corresponding to domain entities are BatchEntity, ExecutionEntity, StepEntity (which is actually part of a JSON field in ExecutionEntity) and OutputFileEntity.

Event handlers

This adapter uses AMQP event handlers reusing the REGARDS IPublisher and IHandler interfaces. It is reusable as long as the mechanisms behind these interfaces are available.

Process management

This adapter uses REGARDS plugins as described below. It is not reusable in a different context.

Workload engine

This adapter uses REGARDS jobs as described below. It is not reusable in a different context.

Order specific design choices

The Processing service has a domain core based on a reactive stack and a clean code architecture. This allows the service to be used in many contexts, however its main use is in conjunction with the Order service, in the REGARDS context, and reusing several REGARDS functionalities. This means that Processing is not used in reactive mode, but in servlet mode, in this context.

OrderProcessInfo, scope and cardinality

A PProcess provides a metadata key/value store, which provides free information to the user. The author of the process can put any useful information for using the process in this metadata.

In the case of processes used by Order, this process info is interpreted as an OrderProcessInfo object, containing:

• limitations on the quantity of data accepted as input,
• the list of required DataType files (for instance, if a process will be applicable only on RAWDATA files, this list contains only RAWDATA),
• the process size forecast, so that Order can pre-calculate an estimate of the size of the output files
• and two informations we will discuss in more details below: scope and cardinality.

Process scope

The scope tells how to launch executions in the context of the same suborder.

The scope can be any of the two following values:

• FEATURE: the process is designed to deal with one feature at a time, and thus there will be one execution per feature in the suborder,
• SUBORDER: the process is designed to deal with several features at a time, and there will be only one execution per suborder, with all the features given as input.

Process cardinality

The cardinality tells how many output files an execution will generate.

The cardinality can be any of the following values:

• ONE_PER_INPUT_FILE: the process is designed to produce one output file for each input file.

  An example of such a process would be a process which produces a companion file for each input file, providing information about the file content, its size, etc.

• ONE_PER_FEATURE: the process is designed to produce one output file for each feature present in the suborder (this has the same result as ONE_PER_INPUT_FILE if there is only one input file per feature, for instance if the process requires only the RAWDATA files and each feature has only one RAWDATA),

  An example of such a process would be to create a compressed archive for each feature, each archive containing all the files associated to the feature.

• ONE_PER_EXECUTION: the process is designed to produce only one output file, no matter what its input is.

  An example of such a process would be to take several images and generate a superposition of all these images, or again producing a compressed archive of all the files present in the input.

Scope/cardinality couples

Since there are 2 scopes and 3 cardinalities, there are 6 possible combinations, but not all of them are meaningful/useful, and there is redundancy.

For instance, the couples:

• scope=FEATURE cardinality=ONE_PER_FEATURE and
• scope=FEATURE cardinality=ONE_PER_EXECUTION

have the same result: the scope says that there is one execution per feature, so they result in the same behaviour.

The other couples are:

• scope=FEATURE cardinality=ONE_PER_INPUT_FILE, which may often also have the same behaviour as the couple above, if there is only one file provided as input for each feature (which may be the case if the process requires only the RAWDATA files as input),
• scope=SUBORDER cardinality=ONE_PER_FEATURE is useful if the process is designed to treat features independently but has better performances when grouping several executions, for instance because it has a big overhead in its initialization,
• scope=SUBORDER cardinality=ONE_PER_INPUT_FILE is useful in the same cases, but now there are several files per feature and they are treated independently,
• scope=SUBORDER cardinality=ONE_PER_EXECUTION is useful if the process is designed to aggregate information from several features ; however there is for now no possibility to group features according to criteria in the same suborder, so the process must accept any group of features.

General algorithm of Order/Processing interaction

Here is a description of the way Order and Processing interact, leaving the details of the Processing inner working out of the picture for more clarity.

Process definitions are REGARDS Plugins

In the context of the Processing service deployed in REGARDS along with the Order service, where processes are used to post-process entity feature files, processes are defined as REGARDS plugins, implementing the IProcessDefinition interface.

These plugins need some configuration:

• allowed dataset for a process
• allowed user role to use a process

This extra information is stored in database as a wrapper for PluginConfiguration named RightsPluginConfiguration. There is a specific REST API for interacting with these wrappers, which are used by the administrator.

Workload engine uses REGARDS Jobs

The provided workload engine uses the REGARDS jobs mechanism. When the workload engine is given an executable, it creates a QUEUED job for execution as soon as possible.

By setting the regards.jobs.pool.size property, we can limit the number of jobs running in parallel, and thus the number of process executions. This limits the overall number of jobs, however, and thus does not permit to limit the number of executions per process.