Skip to content

vincentnam/docker_datalake

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

#f03c15 The project is now hold on 2 differents repository #f03c15

#f03c15 The official main repository is the IRIT Gitlab repository (https://gitlab.irit.fr/datalake/docker_datalake). #f03c15

#f03c15 The secondary repository is the Github initial repository (https://github.com/vincentnam/docker_datalake). #f03c15

The github repository is mirrored to follow the progress of the Gitlab repository and to be synchronized.

Data lake architecture POC hosted on OSIRIM (https://osirim.irit.fr/site/)

This repository is the main working repository for the data lake architecture. The goals are :

  • Create an opensource architecture for data management based on data lake architecture
  • Handle any type of data
  • Handle any volumetry of datas
  • Make it easy to deploy the architecture
  • Make an interoperable architecture through the usage of REST API whenever it is possible

The development of this architecture will integrate a semantic dimension in the development of services and features.

Table of content

Datalake concept

Return to the table of content

Data-oriented project are most of the time driven by the use case. Whether it is for employee data storage, for reporting, for anomaly detection, for application operation, etc.., it requires a specific architecture based on a database. alt text Those kind of architecture needs to be secured, well dimensioned and administered. But if another use case is deployed, the whole process has to be done again for each new use case. It costs time and resources (human and server resources).

The data lake allows to share the whole architecture for security, authentication, server allocation services for any kind of data driven project. Moreover, the solution proposed here allow to integrate any kind of data management, data analysis or reporting solution in a unique solution. alt text

It reduces costs for high volumetry and a high variety of data project or companies. This solution has been designed to integrate solutions that has already been deployed, as authentication systems, database solutions or data analysis solutions. The initial resource investment is higher than simple database solution but it is reduced for new other solutions.

Scientific paper

Return to the table of content

See :

  • "A Zone-Based Data Lake Architecture for IoT, Small and Big Data", Dang, V.N., Zhao, Y., Megdiche, I., Ravat, F., IDEAS 2021 (2021)

The project : context

Return to the table of content

This project is supported by neOCampus, OSIRIM, the CNRS, the IRIT and the SMAC team in IRIT.

This project has been started with a internship for 2nd year of Master in Statistic and Decisional Computing (SID at Université Toulouse 3 - Paul-Sabatier in Toulouse) funded by neOCampus. The project has been continued through a 1year-fixed-term contract by the CNRS and the OSIRIM platform.

  • The IRIT ("Institut de Recherche en Informatique de Toulouse" : https://www.irit.fr/) is a french laboratory under the supervision of the CNRS and the Toulouse universities in the field of computer science.
  • NeOCampus is an operation based in the University with numerous research laboratory. A big part of the network used in this operation is used by sensors and effector. But we can find a lot of other kind of informations. The goal of this project is to create an architecture of data lake that can handle the needs and the data in neOCampus operation.
  • OSIRIM (Observatoire des Systèmes d'Indexation et de Recherche d'Information Multimedia) is one of IRIT's platforms. It is mainly supported by the European Regional Development Fund (ERDF), the French government, the Midi-Pyrénées region and the Centre National de la Recherche Scientifique (CNRS).
  • The SMAC team is interested in modeling and problem solving in complex systems using multi-agen technology. ( https://www.irit.fr/departement/intelligence-collective-interaction/equipe-smac/ )

From March 2020 to now : The project is designed, handled, maintained by DANG Vincent-Nam (for neOCampus / IRIT / OSIRIM / CNRS / Toulouse university).

Beginning 2021 : The project has been joined with a patronage by Modis engineers based on project designed by DANG Vincent-Nam.

The project : design

Return to the table of content

Zones are as in this theoritical solution (see State of art : "Datalake : trends and perspective" paper).

alt text

The architecture design is described in the next diagram :

alt text

Current architecture : Zone-based service-oriented datalake

Return to the table of content

The architecture is divided in 6 zones :

  • Raw data management zone (also known as the landing area)
  • Metadata management zone
  • Process zone
  • Processed data zone (also known as the gold zone)
  • Services zone
  • Security and monitoring zone

Raw data management zone (a.k.a. landing area)

Return to the table of content

The purpose of this area is to handle, store and make available raw data. Each data is stored as is waiting to be processed and transformed into an information.

This area is the core of the architecture, where every data are stored and every service of this architecture are working on. This area is composed by 1 service :

  • Openstack swift (https://wiki.openstack.org/wiki/Swift) :
    • Openstack Swift is a cloud storage software. It is an object-oriented store. It is a part of the Openstack cloud platform deployment and has been built for scale and optimized for durability, availability, and concurrency across the entire data set.
    • Its role is to store all input data as an object. It can handle any type of file or data of any size.

With a network-oriented vision, this zone is a buffer zone.

Metadata management zone

Return to the table of content

-> TODO : redesign this area : - Add metamodel in implementation - Define which metadata to keep in which service (i.e. at the moment, it should be : models (i.e. links betweens metadata in neo4j) and metadata in mongodb) - Seems to be more collections in Mongodb for each type of document

  • MongoDB (https://www.mongodb.com/) :
    • Document oriented NoSQL database. It has been built as a NoSQL database made for high volumetry input.
    • Its role is to store meta data over the data inserted in Openstack Swift and to make it possible to follow and store data over data and be able to know what is stored in Openstack Swift.
    • Operations logs ?

As the Neo4J seems to not be the solution due to legal problem with the neo4j license (only commercial license allow deployment + authentication ), other solution may be evaluated :

Process zone

Return to the table of content

This area is headed by Apache Airflow that will handle and manage every workflow and jobs of data processing and sub services deployed and managed by Apache Airflow :

  • Apache Airflow (https://airflow.apache.org/)

    • Job and workflow scheduler application. This service make it possible to schedule and monitor workflows written in Python.
    • Based on direct acyclic graphs, the data life is easily monitored.
    • It process raw data from the raw data management area to the processed data area by applying custom workflows developed by and for users needs.
  • Apache Spark (https://spark.apache.org/):

    • Used for ressources allocations for data processing on large dataset with high parallelization
    • Used in Lambda or Kappa architecture or real-time data analysis pipeline
    • Warning : some optimizations has to be done to be efficient with Openstack Swift (cf : https://lsds.doc.ic.ac.uk/sites/default/files/swift-analytics_ic2e17_crv.pdf ). In this architecture, we can save file path directly to make it faster in metadata storage (with no overhead as it designed to store this kind of metadata)

The deployment of a Hadoop cluster has been thought but the idea could be not implemented or kept. Indeed, hadoop is designed to handle large files and can't handle well small files (cf default block size : 64MB ou 128MB)

Processed data zone (a.k.a. consumption zone, access zone or gold zone)

Return to the table of content

This area is there to create values over data. Its role is to provide information and allow external application to work on data. The processed data area is supposed to host any database that is needed by users. As no real need have been expressed, no real use case are implemented here. But some use cases have been imagined :

  • InfluxDB (https://www.influxdata.com/):
    • Time serie oriented database with data visualization tools natively integrated.
    • The purpose of this database is to store formatted sensors readings and make it possible to visualize those data easily.
  • Neo4J (https://neo4j.com/):
    • Graph oriented database with data visualization tools natively integrated.
    • The purpose of this database is to store metadata over blob data (image, sound, video) to create a recommendation system for dataset creation.
  • MongoDB (https://www.mongodb.com/):
    • Same database as the metadata database used in the raw data management area.
    • The purpose here is to store data for in-production applications
  • SQL Database (MsSQL Server 2017 ?)

Services zone

Return to the table of content

This functional area includes every service to make this platform user-friendly. At this point (23/11/2020), 3 services have been designed :

  • Data insertion and download services :

    • Composed with 2 services : web GUI and REST API.
      • The web GUI is based on NodeJS server with a React application for user-friendly data insertion. The data are inserted or shown through a graphical user interface to make it easy to use.
      • The REST API is developed with python Flask API to make it possible to programmatically use the data management services (insert data in raw data management area or download data from processed data area)
  • Data stream insertion :

  • Real-time data consumption service :

    • No solution have been found at this point (23/11/2020) to answer this need.
    • The purpose of this service is to make it possible to consume data in a real-time process (initially designed for autOCampus project in neOCampus)
  • Streaming data consumption service :

    • No solution have been found at this point (23/11/2020) to answer this need.
    • The purpose of this service is to make it possible to consume data as a stream for application that works in streaming mode. It has been initialy designed for online machine learning application.

(12/04/2021) Data consumption may be done through RESTful API, direct access to real-time database (InfluxDB as an exemple) or create direct access

API REST, interaction with code

Return to the table of content

Web GUI, interaction with interface

Return to the table of content

(Documentation first draft)

The graphical interface consists of several distinct features. (For the moment, the interface is designed mostly in French)

Upload module :

alt text Files can be uploaded (in "Upload" tab) to the Datalake architecture in 2 ways:

  • Data is pulled from an open server via SSH.
  • Data is pushed directly by the user through the graphical interface. A file splitting is performed if it exceeds for large files (by default at 10 MB with a maximum limit of 5 GB by default (see Openstack Swift Large Object Support)). A drag'n'drop zone allows the user to simply upload his data. It is also possible to select a file via a standard dialog box.

The file type (based on MIME type) must be filled in to assign the appropriate processing pipeline (see Apache Airflow) and an associated metadata model.

alt text The metadata template can be chosen from the list of templates associated with the file type. alt text Once selected, the list of metadata of the model are displayed, allowing to modify the values of these metadata. It is possible to add new/remove metadata or modify the model. The models can be made visible to all users of the architecture or they can be kept confidential.

alt text

Upload tracking module :

A follow-up of the data upload can be done (in the tab "Tracabilité"). It is possible to see the data being uploaded as well as the upload status, the start date and the last update date of the upload status.

alt text

It is possible to view the upload status of the latest data.

alt text

Download module :

The stored data can be downloaded (in the Download tab).

The raw data alt text and processed data are accessible and can be selected for local download.

alt text

Data visualisation module :

A basic visualization of the data inserted in the data lake (currently time series) can be done (in the "Data Visualization" tab). alt text

The data can be viewed in graphical form, allowing for easier selection of the data to be downloaded. They can also be visualized in a table format. alt text

Anomalies module (feature in proof of concept stage) :

A function of detection of anomalies on the times series was set up and makes it possible to visualize on the various measurements the abnormal values of statements of sensors. alt text

Login module :

For the moment, to access the interface's features, it is necessary for a user to authenticate himself. The first access page is the login page which requires a login and password (from Openstack Keystone). alt text

MQTT Flow creation

The management of MQTT sensor retrieval can be managed via the graphical interface allowing to manage the sensor retrievals by MQTT. The flows can be visualized (in the tab "MQTT flows"). It is possible to visualize the MQTT flows in active

alt text as well as all active and non-active flows. alt text A flow can be modified by pressing the "Modify" button. alt text A flow can be disabled to stop listening to data. alt text

User administration

Administrators have additional functionality with user management. It is possible to view the list of users, alt text to see the privileges of a user alt text and add access to a user. alt text

Security, authentication and monitoring zone

Return to the table of content

The purpose of this area is to make it possible to monitor the whole architecture for administrators and give 3 level monitoring. The area has to be adapted to the host platform so services could change with deployment.

  • First monitoring level : User level
    • Openstack Keystone (https://wiki.openstack.org/wiki/Keystone)
      • This service is here to identify and authenticate users in the architecture.
      • It integrate itself well with Openstack Swift and over other authentication services as LDAP or NIS that allow to maintain pre-existing authentication services.
  • Second monitoring level : Service level
  • Third level : Network and system level
    • SNMP + Zabbix :
      • SNMP is a standard protocol of network management. It allow us to follow machine resources consumption and network usage.
      • Combine with Zabbix, this is a service of network and system monitoring with graphical user interface.

This area has to be work more to better design it. Prometheus could be used to monitor Network and System level and other services could be used for other levels.

Data lifecycle

Return to the table of content

This project is fully placed in the Big data world (see "4 V's of Big data", sometimes 5 or more V are described.). We split the data into 2 distinct groups with distinct goals, perspectives and requirements :

  • Batch data
  • Stream data

Batch data main goal is to be stored. Batch data are defined to stay on disk with a long lifespan. Those data will be processed, valuated and consumed on a different timescale than the production.

Stream data main goal is to be processed. Stream data are defined to be generated and as quickly as possible processed for consumption. This data is eventually stored to be consumed again later or in another way.

The architecture handle batch data as main goal but it will be enhanced to handle stream data and eventually near-real time data processing, depending on implementation and tools used.

Main data life cycle : Batch data

Return to the table of content

The global data life in this architecture is described in this diagram :

alt text

  • First, data are inserted simultaneously in Openstack Swift and MongoDB meta database.
  • Then, a webhook is triggered by swift proxy to trigger the Airflow workflow for new data containing metadata over the data inserted.
  • Depending on the data type, the user and the user group / project, the data is processed, transformed and inserted in the corresponding database in the processed data area.
  • Every operations are logged in the metadata database.

Batch data can also be split into 2 distinct groups : light process and heavy process. The light processes include the reading, formatting processes and all the processed made to make the data available. Heavy processes are processes which aim to create data from master data such as data mining, data analysis or machine learning.

alt text

TODO: Explanation of diagram

Light process

Return to the table of content

alt text

alt text

  1. Data are inserted through API designed in the service zone. All batch data are inserted through the RESTful API. For security reasons, the RESTful API (Flask) is accessible behind an Reverse Proxy. The RESTful API handle insertions of data and metadata (respectively Openstack Swift and Mongodb).
  2. Web trigger is raised from the Openstack Swift proxy triggering an Apache Airflow workflow with minimum metadata to retrieve the data.
  3. Apache Airflow workflow process metadata (type, project or every metadata needed to branching through the workflow easily). Data is eventually read from Openstack Swift and each process are logged in metadata zone.
  4. The last step of each workflow has to be the storage of each processed data. Intermediate results could be stored if needed and all the data could even be stored in Openstack Swift.
  5. Once the data processed, it can be consumed through the RESTful API from tools in consumptions zone. A direct access could be designed if needed.

Heavy process

Return to the table of content

alt text

alt text

For heavy process, minor differences can be observed in the process area, otherwise no differences exist.
4. The data is send to Apache Spark cluster. Data are processed in the same way as with Apache Airflow except that resources are allocated by Apache Spark. It can be seen as a tool in Apache Airflow tool box.

Extended data life cycle : Stream data

Return to the table of content

alt text

alt text

alt text Stream data pipeline is different as goals are differents. The first step is to define and instantiate the stream. As it is not a step of data input, it is the 0 step.

  1. User ou administrator has to create a stream. As we can assume that each data in the stream (especially in IoT) as the same metadata, the metadata are defined 1 time for the whole stream. Each sample will be linked to this metadata. This way, Apache Airflow can instantiate the stream through Apache Spark and Apache Spark Stream.

    1. Apache Airflow instantiate a stream lifechecker workflow scheduled at regular interval. It will be able to update metadata (exemple : number of samples).
  2. The sample is send to Apache Spark Streaming. The sample is directly sent to Apache Spark to be processed as faster as possible.

  3. 2 tasks have to be done in parallel. We have to process the data and store it in a consumption tool as the main objective and store it in the raw data storage.

  4. This step is depending on the implementation. Indeed, the more secure way is to define consumption through RESTful API (and reverse proxy) but it could be not possible with tools used and a direct access will be needed.

The same pipeline is defined for near-real time but some customization may be needed. Indeed, the time constraint is conditionned by tools performances and tools used. We place ourselves in near real-time, not in real-time or hard real-time. The goal is to consume the data in the same time scale as the production.

Metadata management

Return to the table of content

alt text

TODO : Explanation - Aimed metadata management system - actul : only mongodb

Metadata management system

Return to the table of content

The metadata is managed in the metadata management system. This system is currently constituted by a NoSQL MongoDB database. This system will be enriched later by a Neo4J database (see What's coming up: metadata management system.

The main interest of the MongoDB database is the possibility to have a semi-structured data model allowing the possibility to modify, add and enrich the metadata document without constraints. There are more than 3000 meta-data models for data description and catalog construction. With the objective of building a data catalog and a data recommendation tool, it is necessary not to restrict to one data format.

Metadata management system : MongoDB

Return to the table of content

MongoDB aims to keep the metadata documents of each data and the various histories. It allows to store via its maximum 16 MB of data a large amount of data. Moreover, the primary objective of the database ' We can store a large volume of data while allowing efficient search in these data, especially via the query tool that MongoDB offers and an implementation of MapReduce.

2 databases in MongoDB have been created for the functioning of the architecture:

  • The "stats" database to keep all the statistics and data needed for the operation of the overall architecture.
  • The "swift" database to keep the metadata associated to each object kept in Openstack Swift
Stats database

The "swift" collection contains 1 unique document instantiated at the initialization of the architecture which evolves throughout the operation of the architecture. This document is built as follows:

{   
    "_id" : MongoDB default ID, 
    "type" : "object_id_file", 
    "object_id" : last_available_swift_id 
}

Its purpose is to hold the identifier by which the next data should be renamed.
Designed as a simple counter at the moment, its purpose is to allow counting of the number of total objects kept.

This collection can be used to keep other data calculated (or not) on the life statistics of the architecture. We can think of : statistics on exchanges, a list of counters or even a list of metadata, search results to accelerate the search of future recommendation tools.

Swift database

The "Swift" database is the metadata database of each data. Each collection corresponds to a user of the database or a container in Swift. Each document is built as follows:

{
 "_id" : MongoDB default ID, 
 "content_type" : type of data / MIME type ,
 "data_processing" : type of data processing in Airflow ("custom" or "default") for pipeline choosing,
 "swift_user" : authenticated user that inserted the data, 
 "swift_container" : Openstack Swift container referring to project / user group,
 "swift_object_id" : id from "object_id_file" in stats database,
 "application" : description of the purpose of the data,
 "original_object_name" : original name ,
 "creation_date" : ISODate("..."),
 "last_modified" : ISODate("..."),
 "successful_operations" : [ ] : list of successful operations done on the data,
 "failed_operations" : [ ] : list of failed operations done on the data, 
 "other_data" : {...} : anything that is needed to know on the data (custom metadata inserted by user) 
}

Needed metadata

Return to the table of content

The required metadata is the metadata that must be filled in to make the architecture work.

Field Utility Filling (automatic / manual) Editable
_id MongoDB default identifier. Allows to uniquely identify the document. Automatic No
content_type Data type Automatic (manual overwrite if needed) Yes
data_processing Type of service to provide "custom" or "default"; a set of default services is set up to provide a minimum service Manual (Default: default) Yes
swift_user User who owns the data, defined as the user who inserts the data Automatic Yes
swift_container swift container in which the data is inserted Manual Yes
swift_object_id Unique identifier for objects in Swift Automatic No
original_object_name Original name of the data; necessary to rename the data in the upload Automatic (manual overwriting possible) Yes
creation_date Date of creation of the object; corresponds to the date of insertion of the data Automatic No

Optional metadata

Return to the table of content

Optional metadata is metadata that has been added and whose absence does not disrupt the proper functioning of the architecture in a basic utilsiation scenario.

Field Utility Filling (automatic / manual) Modifiable
last_modified Last_modified date; allows to follow the life of the data Automatic No
successful_operations List of successful operations on the data; allows to follow the life of the data Automatic No
failed_operations List of failed operations; allows to follow the life of the data Automatic No
application Textual description of the data Manual Yes
Other_data JSON document / Dictionary to add additional metadata in key/value format Manual Yes

The project : features

Return to the table of content

TODO

Virtualization

Return to the table of content

alt text

TODO

Sizing : Kubernetes

Return to the table of content

TODO

Automatic deployment with Ansible

Return to the table of content

The automatic deployment of the architecture is done via Ansible playbooks. For now, only the Docker version installed on Linux servers (Centos 7) has been implemented. It remains to implement the version with containers deployed on a Kubernetes cluster.

To automatically install the architecture, several steps are required:

  • Configure the installation: Configuration files in the folder "./deployment_scripts/ansible/" are to be modified:

    • "./deployment_scripts/ansible/inventories/" contains 2 folders. If it is a production deployment, you have to modify the files in "production", if it is a test deployment, you have to modify the files in "staging".
      • "hosts.yml" contains all the host machines specifications of the services. It is necessary to set up either a password for a user with root access, or an SSH key to put in the "host_vars/.ssh" folder. WARNING: It is better to use SSH keys for security reasons. However, the functionality has not been tested yet.
        • "ansible_ssh_private_key_file" and the path to the corresponding key
        • ansible_sudo_pass" for the root password
    • ./deployment_scripts/ansible/roles" contains the services configuration files. They can be customized as needed.
  • Once the configuration files are updated, the installation is done via this command

    ansible-playbook -i ./deployment_scripts/ansible/inventories/staging/hosts.yml /deployment_scripts/ansible/roles/global/tasks/install_datalake.yml

Security management

Return to the table of content

Security : by design

Return to the table of content

The access to the datalake services is done through 1 single point: the REST API. The only entry is here. Warning: Data streams are done directly by Apache Spark Streaming. However, the implementation of the data streams are done via the REST API. (To work on)

Flow control and networks

Return to the table of content

Still not handled

Security : by tools

Return to the table of content

Section to reformat

Identification : Openstack Keystone

Return to the table of content

Authentication is handled by Openstack Keystone and an LDAP directory. An LDAP directory is deployed by default, but it is possible to configure Openstack Keystone to use a pre-existing authentication base. Calls to the various services require token authentication delivered by Openstack Keystone.

Authentication : Server Kerberos

Return to the table of content

Still not implemented

Monitoring

Return to the table of content

Still not implemented

Monitoring : Kubernetes monitoring tools

Return to the table of content

Still not implemented

Monitoring : low level monitoring : SNMP

Return to the table of content

Still not implemented

Monitoring : Openstack Ceilometer

Return to the table of content

Still not implemented

User spaces compartmentalization

Return to the table of content

Still not implemented

User spaces compartmentalization : Virtualization, Docker and Kubernetes

Return to the table of content

Partially implemented

User spaces compartmentalization : Openstack Keystone

Return to the table of content

TODO

Automation, tests, builds

Return to the table of content

TODO

Automation, tests, builds : CI/CD Pipeline, builds servers

Return to the table of content

Still not implemented

alt text

Automation : Apache Airflow and service lifecheck

Return to the table of content

Still not implemented

Technical description

Return to the table of content

  • MongoDb (Metadata system)
  • Openstack Swift (Raw data storage)
  • Apache Airflow (Process scheduler)
  • Neo4J (Processed data database)
  • Influxdb (Processed data database)
  • NodeJS (Service)
  • React (Service)
  • Flask (Service)
  • Microsoft Server SQL 2017 (Processed data database)

Tool by tool

Return to the table of content

TODO

Airflow

Return to the table of content

Aiflow DAG tools in the apache_airflow/dag/lib folder has a special nomenclature :

  • *tools : all the tools from a database or to process a specific data type
  • *integrator : implemented tool to put data into a specific database

TODO : REWORK THE WHOLE PROCESS DEFINITION

What has been implemented

Return to the table of content

alt text

The project has been tested through a Proof of Concept hosted on Osirim and hosted on several VM on a VM Ware virtualization server. Each service has its own virtual machine except for access zone databases that are all hosted on the same VM. Data storage is made on a NFS bay. Only 2 users are allowed to access this network with SSH. At this point (23/11/2020), the POC is not adapted for this platform and wont be deployed in a production state on OSIRIM.

Network description

Return to the table of content

alt text

Still not implemented TODO : Explanation - 6 differents networks : 1 for each need

alt text

TCP Ports used

Return to the table of content

Raw data area :

  • 8080 : Swift
  • 27017 : MongoDB metadatabase

Process zone :

  • 8081 : Airflow (Webserver)

Consumption zone :

  • Influxdb ports depends on InfluxDB version used (InfluxDB 2 beta version or RC version)

    • 9999 : InfluxDB web interface
    • 8086 : Influxdb web interface
  • 7000 :Neo4J

API description - TODO

Return to the table of content

TODO : Openstack, MongoDB, API Rest for insertion, web gui, etc.. Openstack Swift : REST API (see documentation) MongoDB : API in several languages (Pymongo in Python as an example) Airflow : Web server GUI and REST API (see documentation)

API paths : ...

Data exchanges between services

Return to the table of content

Data exchanges at this point are described in the following schema. alt text TODO : Explanation

Defined

alt text TODO : Explanation

Still undefined

alt text

TODO : Explanation

Data formats

Return to the table of content

Openstack Swift

Return to the table of content

Object inserted in Openstack swift are renamed with a number id. This id is incremented by 1 for every object insert. It allows to follow easily the number of object stored in Openstack Swift.

Only the renamed data are store in Openstack swift. Every metadata are stored in the metadata database (MongoDB). Each object is stored on a container that match to the project or the user group / team.

MongoDB metadata database - TO REFACTOR

Return to the table of content

(23/11/2020) MongoDB contains :

  • "stats" database :
    • "swift" collection :
      • contains only 1 document :
        • { "_id" : ..., "type" : "object_id_file", "object_id" : last_available_swift_id }
        • used to rename and follow Swift object id
      • could be used to store other data
  • "swift" database :
    • a collection for each project has to be created :
      • contains 1 document for each object. Each document contains :
        • "_id" : MongoDB default ID ,
        • "content_type" : type of data / MIME type ,
        • "data_processing" : type of data processing in Airflow ("custom" or "default") for pipeline choosing,
        • "swift_user" : authenticated user that inserted the data,
        • "swift_container" : Openstack Swift container referring to project / user group,
        • "swift_object_id" : id from "object_id_file" in stats database,
        • "application" : description of the purpose of the data,
        • "original_object_name" : original name ,
        • "creation_date" : ISODate("..."),
        • "last_modified" : ISODate("..."),
        • "successful_operations" : [ ] : list of successful operations done on the data,
        • "failed_operations" : [ ] : list of failed operations done on the data,
        • "other_data" : {...} : anything that is needed to know on the data (custom metadata inserted by user)

Getting started

Return to the table of content

Deployment

Return to the table of content

The deployment is done through Ansible playbook in deployment_scripts/ansible.

To deploy the architect, several configuration have to be done :

All have not been tested, but main part of playbook are written. Some modifications may be needed.

Usage

Return to the table of content

Insert a new data

Return to the table of content

TODO : Batch and Stream data insertion

Batch data

To develop a tool to insert data in the datalake, you have to :

  • Get the Swift ID counter and increase it (use "find_one_and_update" to do it with the same operation and reduce chances of data incoherence between 2 instances)
  • Construct metadata to store in mongodb (JSON or Python dictionary)
    • content_type : MIME type of the data
    • swift_user : the swift user that inserted the data
    • swift_container : the swift container where the data is stored
    • swift_object_id : the swift ID given to the data when it has been inserted
    • application : a description of the data usage : the application on which the data is used (optional)
    • swift_object_name : the name of the data (the original name)
    • creation_date : creation date of the document
    • last_modified : date of the last document modification (and get the last operation date)
    • successful_operations : list of successful operations containing (tuple) :
      • execution_date :
      • dag_id : The string of "dag_run" instance containing some informing as the dag_id (ex : <DagRun new_input @ 2020-08-04 14:42:38+00:00: test:tester_neocampus_119:2020-08-04T14:42:38, externally triggered: True>") :
        • If the task has been triggered by the proxy-server trigger, it contains :
          • "manual"
          • the date=
        • If the task has been triggered by the "Check_data_to_process" dag, it contains ("%s_%s_%s:%s"):
          • the user TODO : Openstack, MongoDB, API Rest for insertion, web gui, etc.. Openstack Swift : REST API (see documentation) MongoDB : API in several languages (Pymongo in Python as an example) Airflow : Web server GUI and REST API (see documentation)

          • the container

          • the data swift id

          • the date (datetime)

      • operation_instance : The task id (same format as the dag_id)
    • failed_operations : list of failed operations containing the same as in "successful_operations"
    • other_data : every data that are useful : dictionary
      • For JSON (jsontools.mongodoc_to_influx()):
        • template : dictionary for Influxdb to know what field in the json is for :
          • measurement : str
          • time : datetime
          • fields : list of couple (key:str ,value : Any)
          • tags : list of couple (key:str ,value: Any)
      • For image JPEG / PNG (neo4jintegrator.insert_image()) :
        • image_content : every object that are in the image
          • main_object : the main objects in the image
          • secondary_object : object you can see in image in background
  • Check if the swift container exists. If not, creates it.
  • Put the data in Swift (with content_type and the swift_id)
  • Put the metadata in MongoDB

The "python_test_script.py" is a example script made to add a new data. It has been done to do test but it can be reused to make an insertion script or a REST API. If you want to insert data in the datalake (a file) : use the "insert_datalake()" function in "python_test_script.py"

Stream data

Process a data already inserted

Return to the table of content

There is a document in "stats" database in "swift" collection in MongoDB that contains list of data to process that will be check every 5 minutes by "Check_data_to_process" dag. Adding a swift You'll have to add a document in this list containing :

  • swift_id
  • swift_container
  • swift_user
  • content_type

For each data in this list, it will trigger a "new_input" dag to process this data. DISCLAIMER : "new_input" is actually disable for testing. The actual pipeline is "test" until integration of new pipeline has been done.

Access to services - TODO

Return to the table of content

Create an Airflow job

Return to the table of content

Jobs (or tasks) are done through Operators in Airflow (https://airflow.apache.org/docs/apache-airflow/1.10.14/concepts.html#concepts-operators). The definition of a job or a task is done through a python script. To create a task that will fit in one or more pipeline, an operator has to be used which are defined in the Airflow package. Several operators exists and each one is used for a specific use, including :

  • BashOperator - executes a bash command
  • PythonOperator - calls an arbitrary Python function
  • EmailOperator - sends an email
  • SimpleHttpOperator - sends an HTTP request
  • MySqlOperator, SqliteOperator, PostgresOperator, MsSqlOperator, OracleOperator, JdbcOperator, etc. - executes a SQL command
  • Sensor - an Operator that waits (polls) for a certain time, file, database row, S3 key, etc… In addition to these basic building blocks, there are many more specific operators: DockerOperator, HiveOperator, S3FileTransformOperator, PrestoToMySqlTransfer, SlackAPIOperator… (Cf. Airflow documentation)

The python operator may be the most useful. To use it, there are 2 steps to follow :

  • First, create the python function to be done by the task, example :

def print_context(ds, **kwargs):
print(ds)
return 'Whatever you return gets printed in the logs'
  • Then, define an operator that will run this function. You have to define the DAG (Directed Acyclic Graph) in which the task will be run but each function can be reused in another Operator.

run_this = PythonOperator(
    task_id='print_the_context',
    provide_context=True,
    python_callable=print_context,
    dag=dag,
)

A context can be provided with PythonOperator (always provided in version 2.0) that will allow to give arguments to the function through the **kwargs dict. You can use it as a dictionary and create new keys to provide the data you need. This dictionary contains already a lot of information over the dag run (date, id, etc...) but also contains a "ti" or a "task_instance" (depends on .. ?) key that contains the XCom (for cross-communication) object that allow to pull and push information or objects.

  • You can push data to pass it to another task

kwargs['task_instance'].xcom_push(key="thekey",value="thevalue"))

OR

kwargs["key"] = value
  • You can pull data from previous task from its return or from a xcom_pull()

kwargs['task_instance'].xcom_pull(task_ids='ID_OF_THE_TASK') 

OR

kwargs['ti'].xcom_pull(task_ids='ID_OF_THE_TASK')

Look at the documentation for more information (https://airflow.apache.org/docs/apache-airflow/1.10.14/).

Create an Airflow pipeline

Return to the table of content

Airflow is based on DAG (Directed Acyclic Graph) to implement pipeline / workflows (http://airflow.apache.org/docs/apache-airflow/1.10.14/concepts.html#dags). The definition of a pipeline / workflow is done through a python script The definition of a pipeline is quite straight forward :

  • Define a DAG object

default_args = {
    'start_date': datetime(2016, 1, 1),
    'owner': 'airflow'
}

dag = DAG('my_dag', default_args=default_args)

The DAG can be customized with parameters.

  • Define the relation between task in your dags.

task_1 = DummyOperator('task_1', dag=dag)
task_2 = DummyOperator('task_2', dag=dag)
task_1 >> task_2 # Define dependencies

The "task_2" will be linked to "task_1" and will be run after it. Each task can be define with a run condition (as "all_success", "all_failed", "at least 1 task is successful", etc..). It is possible to create several branches to make several way for processing. The tools used for it are branching operators (see https://airflow.apache.org/docs/apache-airflow/1.10.14/concepts.html#branching) Branching is done the same way but you can link a list of task to branch it. The branching will have to return the task name of the next task to run.

branch_operator >> [way_1 , way_2] 

Integrate a new process pipeline in Airflow

Return to the table of content

04/01/2021 : Right now, it is not possible to easily add a pipeline or a task in Airflow. The way to do it is to change the actual working pipeline. Indeed, only one is triggered by the Openstack Swift proxy when a new data is added.

To add a task or a sub-pipeline / sub-workflow, it will be needed to modify the "./apache_airflow/dags/dag_creator.py" (at the end of the script) and modify the "custom path" in the dag:

custom >> [the_first_task_of_the_sub_pipeline]
the_first_task_of_the_sub_pipeline >> ... >> join

TODO : Add a way to read and parse files in directory and create jobs and dags in function of the content.

For the implementation on OSIRIM, as access are restricted, the best way to add pipeline is to create a python script with :

  • all the functions and the task with operators written in the same file
  • create a dummy DAG (optional)
  • link all the task (eventually with branching)

The new pipeline will be added in the "custom" branch as a new way. Tasks have to be named but 2 tasks have to have different name. The naming convention will be :

PROJECT_USER_TASKNAME

with PROJECT the name of the project or the team in which you work in / with, USER is your username, TASKNAME is a string that quickly describe the task (example : data_cleaning, feature_extraction, etc...). It will be easy and fast to integrate the new pipeline.

What's coming up - Roadmap

Return to the table of content

Incoming : Interoperable datalake federation + distributed metadata management systems

TODO : Implementation

Return to the table of content

TODO list for the project development (i.e. the datalake architecture) TODO : Update TODO list

  • Raw data mangement area :

    • Batch mode for data insertion
      • API Rest
        • Make available the data input
        • API Rest to input 1 or more data
      • Object storage : Openstack swift
        • Unittest : OK
        • Functionaltest : OK
          • Ran: 950 tests in 571.0000 sec.
            • Passed: 883
            • Skipped: 67
            • Expected Fail: 0
            • Unexpected Success: 0
            • Failed: 0
            • Sum of execute time for each test: 502.9311 sec.
        • Trigger for new input to launch a new Airflow job
        • Create middleware for swift proxy (Webhook trigger to launch Airflow jobs)
        • Use X-Webhook in Swift (secure way)
        • Optimizations
  • Metadata management zone

    • MongoDb database for metadata - [ ] Replication for single point of failure problem (REALLY IMPORTANT ! -> if MongoDB datas are corrupted, all data in the datalake are useless)
    • Test other database as metadata management system
    • Implement metadata metamodel
  • Process area

    • Upgrade to version 2.0 (stable) if possible
    • Airflow deployment (docker image)
      • Docker image
      • Installation on VM
      • Resources allocation
        • Parallel executor (Celery)
        • Kubernetes executor
    • Airflow job creation / configuration
      • Handle hook from Swift middleware (Webhook)
      • Set up jobs
    • Find a proper way to add new task / pipeline ( dag from JSON file ?)
  • The processed data area / the gold zone :

    • Relational database (default)
    • Time serie oriented database (visualisation)
      • Json data :
        • Based on templates given (as an input in metadatabase)
    • Document oriented database (transactional vision)
    • Graph database
      • Image files :
        • Jpeg
          • Nodes creation for objects in the file
          • Automatic object detection / segmentation
          • Automatic image analysis with descriptors
      • Recommendation tool
  • The services area :

    • RESTFUL API for data insertion and download
      • Python api with Flask
        • Insertion
        • Download data from database in processed data area
        • Download data from raw data management area
      • Rework with platform update (see Modis branch, projects and issues)
    • Web GUI for data insertion and data visualization
      • Dashboard creation
      • Data download with React + Express backend server
        • Drag'n'drop insertion
        • Progress bar
        • Handle SLO Openstack Swift insertion (Static Large Object) for > 1 Go files
      • Data visualization with React + D3.JS
        • Basic time series visualization
        • Basic graph visualization
        • Complex visualization from several different databases
      • Beautiful dashboard development
    • Real-time data consumption
    • Streaming data consumption
  • Set up a "log" database to log operations on data done

    • Operations are logged in MongoDB MetaDataBase : successful and failed operation (Airflow task + id ) + operations per day
  • Security, Authentication and Monitoring zone - [x] Add Keystone as authentication service

    • Integrate LDAP
    • Integrate NIS
  • Stream mode :

    • Deploy Apache Spark
    • Deploy Apache Spark Stream
    • Add stream creation in REST API
  • AutomaticDeployment : Docker + Kubernetes + Ansible

    • Docker
      • MongoDB enterprise container
      • Openstack Swift container
      • Apache Airflow container (Official container)
      • Apache Spark
      • Processed data zone container (InfluxDB, SQL database)
      • (NEEDED : Security design and implementation) Openstack Keystone container
      • (NEEDED : Process zone design and rework : add Apache Spark) Apache Spark container
    • Ansible
      • Create Ansible playbook for core services
        • Test whole automatic deployment
      • Create playbook for whole architecture
        • Add GUI deployment
        • Add Apache Spark
        • Add processed data zone (for databases)
        • Add multi-node deployment
    • Kubernetes deployment

TODO : Design

Return to the table of content

  • Add metadata over transformed data in the goldzone (and be able to find the list of process done to create this processed data)

  • Design metadata model

    • Data lifecycle metadata
    • Design metadata model for data
      • Data type
    • Redesign logs data
    • Stream metamodel

Another approach to metadata management will be adopted with distributed metadata management.

  • Security and monitoring area (REQUIRED : AutomaticDeployment )

    • Design security services
    • Design monitoring services :
      • Low level monitoring (network)
      • Mid level monitoring (process)
      • High level monitoring (user consumption)
  • API REST

    • Add new job
    • Create new stream
  • Streaming mode for data insertion

    • Stream creation handling

TODO : Documentation

Return to the table of content

  • Set SSPL license
  • Fill empty sections

Other information

Return to the table of content

Problems already encountered

Return to the table of content

MsSQL 20xx

Return to the table of content

The Microsoft Server SQL 20xx (i.e. 2017 or 2019) are deployed through Docker container. A password is needed to set up the database which come from :

  • Environment with a variable
  • Parameter given to the container when the container is ran
  • A hidden file ".env" in the host folder containing environment variables

If this password is not set, the container will crash on boot.

Airflow

Dont name your task the same name of the callable (python function) : it will lead to an error

Tools not used

Return to the table of content

In raw data zone

Return to the table of content

  • HBase : need for raw input data, HBase would have been used as a key / value database while it's a column store database + difficult to handle raw data reading
  • Apache Nifi : not usefull and doesn't fit in the architecture and replaced with Apache Spark with Apache Spark Stream and MQTT
  • Apache Kafka : As MQTT and Apache Spark Stream are / will be used, Apache Kafka isn't needed anymore.

Return to the table of content

  • Talend : difficulties to install on Linux + difficulties to find version that can be integrated in the POC

In processed data zone

Return to the table of content

Nothing at this moment.

More documentation

Return to the table of content

Other markdown files are in folder of each service containing some more information over the service. A pdf is available in the repository. This pdf contains the internship report that I made for the internship. It is mainly made of design thinking.

Tools used in this architecture also have documentation :

Versioning Rules

Return to the table of content

Versions are defined like this : X.Y.Z-a X : Major (integer) Y : Minor (integer) Z : Build (integer) a : Patches (letter)

08/06/2021 : Major : 1.0.0 will be released with a production-ready solution with tests :

  • Security is enabled with Openstack Keystone
  • Automatic deployment, Docker
  • First version of web GUI
  • Minimum service with processes and Apache Spark are deployed
  • Stream data handled
  • Metadata management system with MongoDB
  • Monitoring tools deployed for administration at least (optional)

2.0.0 will be focused on metadata management and Kubernetes. (TODO)

Minor : each minor are released when a project is ended :

  • Minimum service in Apache Airflow for some data types (STARTED)
  • Apache Spark and Apache Spark Stream with integration with Apache Airflow
  • Web GUI first version with insertion, download and first data visualization tools (STARTED)
  • Openstack Keystone and integration in other services
  • Metadata management system improvement (STARTED)
  • Monitoring tools (optional)

Build : each build define a sprint or a (bunch of) feature(s) developped.

Patches : for a bug fix

License

Return to the table of content

The whole project is (C) 2020 March, DANG Vincent-Nam [email protected]

during an end of study internship for Université ToulAPI description - TODOouse 3 Paul-Sabatier (FRANCE), IRIT and SMAC Team, neOCampus as original designer of the project / solution / architecture and basis code owner

and for CNRS (as a 1-year fixed-term contract) as developer of the project / solution / architecture and

is licensed under the SSPL, see `https://www.mongodb.com/licensing/server-side-public-license'.

Institutions involved


Scientific interest group neOCampus (funder of the internship at the initiative of the project):

A group of laboratories, organizations and industrialists to build the campus of the future

https://www.irit.fr/neocampus/fr/

alt text


Toulouse University : Université Toulouse 3 - Paul Sabatier

https://www.univ-tlse3.fr/

alt text


Toulouse computer science research laboratory :

IRIT - Insitut de Recherche en Informatique de Toulouse

https://www.irit.fr/

alt text


French National Center for Scientific Research - Centre national de la recherche scientifique

http://www.cnrs.fr/en

alt text


Engineers from Modis France (https://www.modisfrance.fr/)

alt text

worked on this project in a patronage project with Université Toulouse 3 Paul-Sabatier and neOCampus.

Contacts

Return to the table of content

04/01/2021 :

*DANG Vincent-Nam (Repository basis code owner, intern and project main maintainer)

[email protected] / [email protected]

https://www.linkedin.com/in/vincent-nam-dang/