Jumia is a know-how firm born in 2012, current in 14 African international locations, with its essential headquarters in Lagos, Nigeria. Jumia is constructed round a market, a logistics service, and a cost service. The logistics service allows the supply of packages by way of a community of native companions, and the cost service facilitates the funds of on-line transactions inside Jumia’s ecosystem. Jumia is current in NYSE and has a market cap of $554 million.
On this put up, we share a part of the journey that Jumia took with AWS Skilled Providers to modernize its information platform that ran underneath a Hadoop distribution to AWS serverless based mostly options. Among the challenges that motivated the modernization have been the excessive price of upkeep, lack of agility to scale computing at particular occasions, job queuing, lack of innovation when it got here to buying extra fashionable applied sciences, complicated automation of the infrastructure and purposes, and the lack to develop regionally.
Answer overview
The fundamental idea of the modernization undertaking is to create metadata-driven frameworks, that are reusable, scalable, and in a position to answer the completely different phases of the modernization course of. These phases are: information orchestration, information migration, information ingestion, information processing, and information upkeep.
This standardization for every part was thought-about as a option to streamline the event workflows and decrease the chance of errors that may come up from utilizing disparate strategies. This additionally enabled migration of various sorts of information following an analogous strategy whatever the use case. By adopting this strategy, the information dealing with is constant, extra environment friendly, and extra simple to handle throughout completely different initiatives and groups. As well as, though the use circumstances have autonomy of their area from a governance perspective, on high of them is a centralized governance mannequin that defines the entry management within the shared architectural elements. Importantly, this implementation emphasizes information safety by imposing encryption throughout all companies, together with Amazon Easy Storage Service (Amazon S3) and Amazon DynamoDB. Moreover, it adheres to the precept of least privilege, thereby enhancing total system safety and decreasing potential vulnerabilities.
The next diagram describes the frameworks that have been created. On this design, the workloads within the new information platform are divided by use case. Every use case requires the creation of a set of YAML recordsdata for every part, from information migration to information movement orchestration, and they’re mainly the enter of the system. The output is a set of DAGs that run the precise duties.
Within the following sections, we focus on the goals, implementation, and learnings of every part in additional element.
Knowledge orchestration
The target of this part is to construct a metadata-driven framework to orchestrate the information flows alongside the entire modernization course of. The orchestration framework supplies a sturdy and scalable answer that has the next capacities: dynamically create DAGs, combine natively with non-AWS companies, enable the creation of dependencies based mostly on previous executions, and add an accessible metadata technology per every execution. Due to this fact, it was determined to make use of Amazon Managed Workflows for Apache Airflow (Amazon MWAA), which, by way of the Apache Airflow engine, supplies these functionalities whereas abstracting customers from the administration operation.
The next is the outline of the metadata recordsdata which are offered as a part of the information orchestration part for a given use case that performs the information processing utilizing Spark on Amazon EMR Serverless:
proprietor: # Use case proprietor
dags: # Checklist of DAGs to be created for this use case
- identify: # Use case identify
sort: # Kind of DAG (may very well be migration, ingestion, transformation or upkeep)
tags: # Checklist of TAGs
notification: # Defines notificacions for this DAGs
on_success_callback: true
on_failure_callback: true
spark: # Spark job data
entrypoint: # Spark script
arguments: # Arguments required by the Spark script
spark_submit_parameters: # Spark submit parameters.
The thought behind all of the frameworks is to construct reusable artifacts that allow the event groups to speed up their work whereas offering reliability. On this case, the framework supplies the capabilities to create DAG objects inside Amazon MWAA based mostly on configuration recordsdata (YAML recordsdata).
This explicit framework is constructed on layers that add completely different functionalities to the ultimate DAG:
- DAGs – The DAGs are constructed based mostly on the metadata data offered to the framework. The information engineers don’t have to jot down Python code with a purpose to create the DAGs, they’re mechanically created and this module is in command of performing this dynamic creation of DAGs.
- Validations – This layer handles YAML file validation with a purpose to forestall corrupted recordsdata from affecting the creation of different DAGs.
- Dependencies – This layer handles dependencies amongst completely different DAGs with a purpose to deal with complicated interconnections.
- Notifications – This layer handles the kind of notifications and alerts which are a part of the workflows.

One facet to think about when utilizing Amazon MWAA is that, being a managed service, it requires some upkeep from the customers, and it’s essential to have a great understanding of the variety of DAGs and processes that you just’re anticipated to have with a purpose to fine-tune the occasion and acquire the specified efficiency. Among the parameters that have been fine-tuned in the course of the engagement have been core.dagbag_import_timeout, core.dag_file_processor_timeout, core.min_serialized_dag_update_interval, core.min_serialized_dag_fetch_interval, scheduler.min_file_process_interval, scheduler.max_dagruns_to_create_per_loop, scheduler.processor_poll_interval, scheduler.dag_dir_list_interval, and celery.worker_autoscale.
One of many layers described within the previous diagram corresponds to validation. This was an essential part for the creation of dynamic DAGs. As a result of the enter to the framework consists of YML recordsdata, it was determined to filter out corrupted recordsdata earlier than making an attempt to create the DAG objects. Following this strategy, Jumia might keep away from undesired interruptions of the entire course of. The module that really builds DAGs solely receives configuration recordsdata that observe the required specs to efficiently create them. In case of corrupted recordsdata, data concerning the precise points is logged into Amazon CloudWatch in order that builders can repair them.
Knowledge migration
The target of this part is to construct a metadata-driven framework for migrating information from HDFS to Amazon S3 with Apache Iceberg storage format, which includes the least operational overhead, supplies scalability capability throughout peak hours, and ensures information integrity and confidentiality.
The next diagram illustrates the structure.

Throughout this part, a metadata-driven framework in-built PySpark receives a configuration file as enter in order that some migration duties can run in an Amazon EMR Serverless job. This job makes use of the PySpark framework because the script location. Then the orchestration framework described beforehand is used to create a migration DAG that runs the next duties:
- The primary process creates the DDLs in Iceberg format within the AWS Glue Knowledge Catalog utilizing the migration framework inside an Amazon EMR Serverless job.
- After the tables are created, the second process transfers HDFS information to a touchdown bucket in Amazon S3 utilizing AWS DataSync to sync buyer information. This course of brings information from all of the completely different layers of the information lake.
- When this course of is full, a 3rd process converts information to Iceberg format from the touchdown bucket to the vacation spot bucket (uncooked, course of, or analytics) utilizing once more another choice of the migration framework embedded in an Amazon EMR Serverless job.
Knowledge switch efficiency is healthier when the dimensions of the recordsdata to be transferred is round 128–256 MB, so it’s advisable to compress the recordsdata on the supply. By decreasing the variety of recordsdata, metadata evaluation and integrity phases are lowered, rushing up the migration part.
Knowledge ingestion
The target of this part is to implement one other framework based mostly on metadata that responds to the 2 information ingestion fashions. A batch mode is accountable for extracting information from completely different information sources (corresponding to Oracle or PostgreSQL) and a micro-batch-based mode extracts information from a Kafka cluster that, based mostly on configuration parameters, has the capability to run native streams in streaming.
The next diagram illustrates the structure for the batch and micro-batch and streaming strategy.

Throughout this part, a metadata-driven framework builds the logic to carry information from Kafka, databases, or exterior companies, that shall be run utilizing an ingestion DAG deployed in Amazon MWAA.
Spark Structured Streaming was used to ingest information from Kafka matters. The framework receives configuration recordsdata in YAML format that point out which matters to learn, what extraction processes ought to be carried out, whether or not it ought to be learn in streaming or micro-batch, and wherein vacation spot desk the data ought to be saved, amongst different configurations.
For batch ingestion, a metadata-driven framework written in Pyspark was carried out. In the identical method because the earlier one, the framework acquired a configuration in YAML format with the tables to be migrated and their vacation spot.
One of many points to think about in one of these migration is the synchronization of information from the ingestion part and the migration part, in order that there isn’t a lack of information and that information isn’t reprocessed unnecessarily. To this finish, an answer has been carried out that saves the timestamps of the final historic information (per desk) migrated in a DynamoDB desk. Each forms of frameworks are programmed to make use of this information the primary time they’re run. For micro-batching use circumstances, which use Spark Structured Streaming, Kafka information is learn by assigning the worth saved in DynamoDB to the startingTimeStamp parameter. For all different executions, precedence shall be given to the metadata within the checkpoint folder. This manner, you can also make positive ingestion is synchronized with the information migration.
Knowledge processing
The target on this part was to have the ability to deal with updates and deletions of information in an object-oriented file system, so Iceberg is a key answer that was adopted all through the undertaking as delta lake recordsdata due to its ACID capabilities. Though all phases use Iceberg as delta recordsdata, the processing part makes intensive use of Iceberg’s capabilities to do incremental processing of information, creating the processing layer utilizing UPSERT utilizing Iceberg’s capability to run MERGE INTO instructions.
The next diagram illustrates the structure.

The structure is much like the ingestion part, with simply adjustments to the information supply to be Amazon S3. This strategy quickens the supply part and maintains high quality with a production-ready answer.
By default, Amazon EMR Serverless has the spark.dynamicAllocation.enabled parameter set to True. This selection scales up or down the variety of executors registered inside the software, based mostly on the workload. This brings a whole lot of benefits when coping with several types of workloads, however it additionally brings issues when utilizing Iceberg tables. As an illustration, whereas writing information into an Iceberg desk, the Amazon EMR Serverless software can use a lot of executors with a purpose to velocity up the duty. This can lead to reaching Amazon S3 limits, particularly the variety of requests per second per prefix. For that reason, it’s essential to use good information partitioning practices.
One other essential facet to think about in these circumstances is the article storage file structure. By default, Iceberg makes use of the Hive storage structure, however it may be set to make use of ObjectStoreLocationProvider. By setting this property, a deterministic hash is generated for every file, with a hash appended straight after write.information.path. This may significantly decrease throttle requests based mostly on object prefix, in addition to maximize throughput for Amazon S3 associated I/O operations, as a result of the recordsdata written are equally distributed throughout a number of prefixes.
Knowledge upkeep
When working with information lake desk codecs corresponding to Iceberg, it’s important to interact in routine upkeep duties to optimize desk metadata file administration, stopping a lot of pointless recordsdata from accumulating and promptly eradicating any unused recordsdata. The target of this part was to construct one other framework that may carry out all these duties on the tables inside the information lake.
The next diagram illustrates the structure.

The framework, in addition to the opposite ones, receives a configuration file (YAML recordsdata) indicating the tables and the checklist of upkeep duties with their respective parameters. It was constructed on PySpark in order that it might run as an Amazon EMR Serverless job and may very well be orchestrated utilizing the orchestration framework similar to the opposite frameworks constructed as a part of this answer.
The next upkeep duties are supported by the framework:
- Expire snapshots – Snapshots can be utilized for rollback operations in addition to time touring queries. Nonetheless, they’ll accumulate over time and might result in efficiency degradation. It’s extremely advisable to recurrently expire snapshots which are not wanted.
- Take away previous metadata recordsdata – Metadata recordsdata can accumulate over time similar to snapshots. Eradicating them recurrently can also be advisable, particularly when coping with streaming or micro-batching operations, which was one of many circumstances of the general answer.
- Compact recordsdata – Because the variety of information recordsdata will increase, the variety of metadata saved within the manifest recordsdata additionally will increase, and small information recordsdata can result in much less environment friendly queries. As a result of this answer makes use of a streaming and micro-batching software writing into Iceberg tables, the dimensions of the recordsdata tends to be small. For that reason, a technique to compact recordsdata was crucial to reinforce the general efficiency.
- Arduous delete information – One of many necessities was to have the ability to carry out arduous deletes within the information older than a sure time period. This suggests eradicating expiring snapshots and eradicating metadata recordsdata.
The upkeep duties have been scheduled with completely different frequencies relying on the use case and the precise process. For that reason, the schedule data for this duties is outlined in every of the YAML recordsdata of the precise use case.
On the time this framework was carried out, there was no any computerized upkeep answer on high of Iceberg tables. At AWS re:Invent 2024, Amazon S3 Tables performance has been launched to automatize the upkeep of Iceberg Tables . This performance automates file compaction, snapshot administration, and unreferenced file elimination.
Conclusion
Constructing a knowledge platform on high of standarized frameworks that use metadata for various points of the information dealing with course of, from information migration and ingestion to orchestration, enhances the visibility and management over every of the phases and considerably quickens implementation and growth processes. Moreover, by utilizing companies corresponding to Amazon EMR Serverless and DynamoDB, you’ll be able to carry all the advantages of serverless architectures, together with scalability, simplicity, versatile integration, improved reliability, and cost-efficiency.
With this structure, Jumia was in a position to scale back their information lake price by 50%. Moreover, with this strategy, information and DevOps groups have been in a position to deploy full infrastructures and information processing capabilities by creating metadata recordsdata together with Spark SQL recordsdata. This strategy has lowered turnaround time to manufacturing and lowered failure charges. Moreover, AWS Lake Formation offered the capabilities to collaborate and govern datasets on numerous storage layers on the AWS platform and externally.
Leveraging AWS for our information platform has not solely optimized and lowered our infrastructure prices but in addition standardized our workflows and methods of working throughout information groups and established a extra reliable single supply of reality for our information property. This transformation has boosted our effectivity and agility, enabling quicker insights and enhancing the general worth of our information platform.
– Hélder Russa, Head of Knowledge Engineering at Jumia Group.
Take step one in the direction of streamlining the information migration course of now, with AWS.
In regards to the Authors
Ramón Díez is a Senior Buyer Supply Architect at Amazon Internet Providers. He led the undertaking with the agency conviction of utilizing know-how in service of the enterprise.
Paula Marenco is a Knowledge Architect at Amazon Internet Providers, she enjoys designing analytical options that carry gentle into complexity, turning intricate information processes into clear and actionable insights. Her work focuses on making information extra accessible and impactful for decision-making.
Hélder Russa is the Head of Knowledge Engineering at Jumia Group, contributing to the technique definition, design, and implementation of a number of Jumia information platforms that help the general decision-making course of, in addition to operational options, information science initiatives, and real-time analytics.
Pedro Gonçalves is a Principal Knowledge Engineer at Jumia Group, accountable for designing and overseeing the information structure, emphasizing on AWS Platform and datalakehouse applied sciences to make sure strong and agile information options and analytics capabilities.