Big data marketing. The story of how technology behind Allegro marketing works.
Marketing is a very important department in every company. In case of Allegro, marketing is especially difficult because you have so many products to promote. In this post we will tell the story of a platform we built for marketing purposes.
Background #
Allegro is the biggest e-commerce platform in Poland and one of top 10 largest e-commerce platforms worldwide. Our catalog holds almost 200 million offers at this moment (December 2020), and the number is still growing. The marketing team uses tools such as Google Merchant Center and Facebook Ads in order to advertise Allegro offers and get more traffic to the platform. To integrate with them we need to prepare an XML file containing information about our offers. Such XML files are called “feed” . We will use this name later on.
Since we need to find specific offers, using a search engine may seem natural. However, Allegro’s search engine is not suited for this task. Feed could contain millions of offers, what would result in deep pagination issues. A decision was made to simply generate static files using batch jobs run in our Hadoop ecosystem. The ability to handle large volumes of data, powerful query capabilities as well as access to various datasets across the platform were major advantages. Apache Spark, an already tried and tested tool, was an obvious choice.
Since we didn’t expect the number of feeds to exceed a few dozen, every
feed was calculated in a separate (but executed in parallel) Spark job.
Business users created every feed definition by providing predicates that offers must
satisfy to be included in feed, as well as expected XML format and recipient.
You can see that architecture in the diagram below. AggregateGeneratorJob and
FeedGeneratorJob were batch jobs written in Apache Spark. First one collected
data from different sources on Hive and HDFS, then assembled them into a single
Parquet-based file called simply “aggregate” (we will use this name later on).
Second job, FeedGeneratorJob generated and uploaded a single feed (XML file)
to S3. All jobs were run in parallel.
But soon, against initial assumptions, a number of feeds exploded. Eventually, we encountered as much as… 1300 feeds! Updating all of them, to present current data in advertisements, took more than 24 hours. We managed to improve this situation a little by vertical scaling and removing some of unused/poor performing feeds. However, it was just a temporary improvement, since it still took as much as 13 hours to refresh all the feeds.
We were yet to find out that poor performance was just the tip of the iceberg. Much bigger problem was the architecture that no longer suited our needs and made implementing new features time-consuming. Codebase used a then acclaimed cake (anti)pattern that turned out to work poorly in connection with Spark. It caused serious serialization issues. Add to that leaky monitoring and handwritten scheduler, and you will get a full picture of our despair. Besides, the tool itself became very important. It handled more and more integrations and was crucial for the company.
Brave new world solution #
At that moment we knew that we needed a new solution. We decided that our target solution should let us:
- Reduce the execution time to 1h (2h at most) while keeping same amount of resources
- Query over any offer property (in old solution we had only some predefined predicates)
- Choose offer by a key (or in programmers language: group by + aggregate)
- Introduce new data sources quickly
- Create feed with arbitrary size: from just a few offers to whole catalog
- Scale horizontally
- Last but not least: integrate with our partners not only by files but also using an event-based approach (streaming API)
These requirements would be easy to comply with in case of a “normal” shop with a few thousands products. However, Allegro operates on a much larger scale of:
- almost 200M offers (and still growing)
- ~25-40M changes in offers per day
Also, Allegro is based on microservices architecture and we don’t have a single DB with full information about the system. This leaves us with yet another problem: how to gather all needed data. We have to use information on offers, sellers, campaigns, ratings, products and few others. So the first item on our TODO list was to find a solution for collecting the data. In Allegro most of the services use Hermes as a message broker. Also, all of the data that is sent by Hermes is dumped to HDFS in near real-time manner. To make this clearer, let me show you that on diagram:
At that moment, we wondered which approach would suit our requirements best. We saw three options here:
- Find some existing solution in our platform and customize it for our needs
- Use Hermes topics directly (online)
- Collect data from HDFS (offline)
First option would be nice, but there was one problem… we haven’t found any suitable source. So basically, we had to choose between collecting all data online vs offline. Beside the most obvious difference, latency, what else differentiates these solutions?
It is always more difficult to join data online. We would need to maintain a database with the whole state and we would be prone to all kinds of concurrency-related bugs. In case of any detected problem we would have to recover using historical data.
Offline solution would be similar to what we had in the old platform
(AggregateGeneratorJob described before). Joins between various data sources
would be straightforward. We wouldn’t have any problems with concurrency.
Recreating data is easy, since basically it is done on every job execution,
although we pay for that with latency. The question though was how long would it take
to create such aggregate and how much latency we would get at that stage.
Considering it was easy to implement we decided to simply measure it. In the end it
turned out not that bad: in typical cases we were able to maintain
latency of about 30 minutes.
That was acceptable for a start. In case of it being not enough, we could always transform it later into delta architecture and read the newest data (or at least some subset of it, for example products prices) from Hermes to bring it up-to-date.
Once we had a data source, we had to find a way to generate feeds based on it. We were a bit biased against Apache Spark, because of poor performance and hard maintainability of the old platform. Back then we didn’t know that problem was in our solution, not in Spark itself. That’s why we decided to spend a few weeks on research. We made a few prototypes based on Spark Streaming, Kafka Streams and on databases. We even had an idea of writing our own engine for computation. During that research we came up with the idea of generating feeds in an efficient way and… we realized that it will be pretty easy to implement in Spark! We also made an important decision: we will focus on generating files, and get back to streaming API later.
I am speed #
Basically, in order to generate feed we need to look through offers catalog and find all offers matching defined criteria. In a database you typically use indexes on a subset of fields to simplify searching. Since we need the possibility of making predicates on all fields as well as of integrating all offers with our partner, we decided to go for linear scanning. Is it bad? Well, it depends on next steps. Since we decided on linear scanning, we knew that complexity of our process would be at least O(N). We could handle that, but only as long as we would be able to make complexity independent of the number of feeds (integrations). Even more importantly, we had to take care of scalability. It would be best to partition data and calculate it concurrently, while sharing as little common state as possible.
In the old process every feed was calculated by a separate job. Although we had hundreds of integrations, lots of them use the same XML schema (later called “template”). Also, lots of feeds use similar predicates, so their results can be cached and reused. Therefore, our first priority was to calculate everything in one go. In order to achieve that we simply divide our aggregate into partitions and for each partition we evaluate predicates to figure out which feed includes which offer. When we know that, we also know in which templates we need to render an offer. At the end of the process we pivot and write data to appropriate partition-files on S3. From S3 we serve partitioned files using our file server that knows how to assemble parts and serve them as a single piece.
Ok, so how much speedup we gained thanks to that approach? After rewriting we were able to recalculate all feeds in a little over 1h (comparing to 13h previously). It wasn’t all, though. Not only have we sped the process up 13 times, we also reduced memory usage twofold! And well, in the end we used the same tools, but in a better way.
Streaming API #
After we drank our champagne, and cooled down a little bit, we had to return to the problem with providing streaming API. Following ambitious targets of our marketing business, we wanted to integrate Allegro’s offers in a more effective way. This type of integration results in smaller latency of products’ updates and also fewer resources are required on the partner’s side.
At that moment we returned to the problem stated before: what should we use as a source of data? Catching every event in the moment that it was produced would be very difficult due to the scale and resources required to handle such traffic.
Moreover, being a constant listener to all events emitted in our platform and sending them instantly to various partners’ APIs brings no benefits in terms of data freshness. This is due to the fact that updates are not applied immediately by partners’ sides - even though the latency is lower than in the XML solution, it still occurs and can take up to a couple of hours.
We decided that we can start with taking data from offers’ aggregate built for file-based feeds. We didn’t want to send all offers that should be integrated with a partner at every job’s run because in most cases we would generate redundant events. Some offers didn’t change between two successive runs at all, some of them were newly added or removed from the feed, but in most cases they had only partial updates e.g. price change. So we had the idea of sending just the difference between the previous and current event feed state. How? Here’s a simplified version of algorithm for this approach:
- load latest state of all Allegro’s offers from HDFS - let’s call it aggregate,
- extract from aggregate and save on HDFS only the offers that should be included in event feed - let’s call it snapshot X,
- load the previous state of the event feed (from the previous run) - snapshot Y,
- make a full join on X and Y using offer’s unique key - dataset Z of type
Tuple(OfferStateX, OfferStateY), - decide to generate appropriate events based on dataset Z:
- if both values are non-empty, generate an event with the calculated difference between state X and Y,
- if the value of X is empty, generate an event on removal from the feed,
- if the Y value is empty, generate an event on addition of a new offer to the event feed,
- send generated events to Kafka topic that is constantly consumed by the service (connector) responsible for sending offers to a marketing partner,
- save snapshot X on HDFS (in the next run it will act as a snapshot Y)
I’m sure you’re wondering how much latency this solution adds. Well, it turned out to be only 20 minutes and in our case it is totally acceptable. It is also worth mentioning that our Kafka topic is scalable in case a new partnership appears. This is because the event model contains information about its destinations. Thanks to this approach, we reduce the amount of data sent, thus limiting the traffic of millions of sent events to just tens of thousands.
Self-healing system #
Every complex system is prone to inconsistencies. Especially when this complexity increases and it is hard to stop that - as it was before our big refactor. New architecture let us create a self-healing and fully controllable system that is convenient to maintain even taking into account the scale we face everyday. When designing the architecture we focused mainly on two things: availability and control.
Availability #
The first step that should be considered is: how to properly define the responsibilities of individual system components?
System works as a sum of cooperating elements. But it is easier to maintain these elements when they have very specific tasks to handle. There are three main components in our system (introduced earlier):
- Aggregator - knows everything about every offer’s current state. It gathers all needed data and saves them to HDFS,
- Generator - it takes data generated by Aggregate, filters it and prepares for delivery to a partner,
- Connector (possibility of having many connectors) - holds integration with partner, acts as a data mapper and sender.
Aggregator and Generator are based on a complementary set of information about the whole system and offers’ state at a certain point in time. So in case the aggregate contains damaged offers and the generator already took it to prepare for sending, in the next cycle it will get overwritten by fixed ones. This happens because each cycle run’s results overwrite the previous ones. Additionally, both Aggregate and Generator stage results are persisted on HDFS. Thanks to this we can run the whole computation for any period of time and go back to any system state. Also, the Generator stage can be based on data generated at any time. In case Aggregate is failing while generating new data, Generator works properly using earlier data.
Then, we have a Connector. It consumes events from Kafka and pushes them, in appropriate form, on partner’s API. It has no responsibility for checking data or state correctness. It simply gets what the Generator prepared and tries to deliver it to a partner. Thanks to this separation of responsibility, Connector is not dependent on Generator - even if Generator has a breakdown, the Connector at worst may have nothing to do.
Control #
In the previous paragraph we mentioned a few processing issues we are struggling with. However, we also proved that despite this our system can still work in such conditions - maybe not as effectively as in standard scenarios, but it still does. To react faster, we’ve managed to make quite “garbage data”-resistant, notifications-based alerting system that will alarm about anomalies occuring during computation. In short, if the difference between states of previous and current Generator run is significant (experience based numbers), the system will stop and inform us about it so that we can decide if this change is acceptable or not. (By difference between states I mean difference between parameters such as feed’s offer count, number of offers’ parameters changes etc.) Once the change is approved, the system returns to its work. Otherwise, data is not propagated from Generator to Kafka, resulting in lack of data to be consumed by Connector. Even if we pass some incorrect data to a partner and it will be too late to retreat, we have a special mechanism refreshing any offer that was updated more than 28 days ago. So if an offer wasn’t updated for such a long time, it doesn’t matter if it is damaged or not – it will be refreshed eventually.
Summary #
Key takeaway points:
- Just because something does not work well it doesn’t mean the tool is bad. Maybe there is something wrong with the way you are using it?
- Ideas can be complex but it doesn’t mean that they have to be complicated!
- Research is key. Even if your business tells you there is no time for it, insist on it. Otherwise you will end up spending even more time on fixes.
- Apache Spark is a beast. It can simplify your computation dramatically and give amazing results with it, but at the same time you need to think more about how your data will be calculated. One small problem may result in slow computation. Unfortunately lots of them are hard to notice.
- Join us if you like such challenges :-)