This post was co-written by Mahesh Pasupuleti and Gaurav Shah from Poshmark.
Poshmark is a leading social marketplace for new and secondhand styles for women, men, kids, pets, home, and more. By combining the human connection of physical shopping with the scale, ease, and selection benefits of Ecommerce, Poshmark makes buying and selling simple, social, and sustainable. Its community of more than 80 million registered users across the US, Canada, Australia, and India is driving a more sustainable future for the fashion industry.
An important goal to achieve for any organization is to grow the top line revenue. Top line revenue refers to the total value of sales of an organization’s services or products. The two main approaches organizations employ to increase revenue are to expand geographically to enter new markets and to increase market share within a market by improving customer experience (CX).
Improving CX is a well-known guideline to attract and retain customers and thereby increase the market share. In this post, we share how Poshmark improved CX and accelerated revenue growth by using a real-time analytics solution. We discuss how to create such a solution using Amazon Kinesis Data Streams, Amazon Managed Streaming for Kafka (Amazon MSK), Amazon Kinesis Data Analytics for Apache Flink; the design decisions that went into the architecture; and the observed business benefits by Poshmark.
High-level challenge: The need for real-time analytics
Previous efforts at Poshmark for improving CX through analytics were based on batch processing of analytics data and using it on a daily basis to improve CX. Although these batch analytics-based efforts were successful to some extent, they saw opportunities to improve the customer experience with real-time personalization and security guidance during the customer’s interaction with the Poshmark app. The customer insights gathered from the batch analytics couldn’t be paired with the current customer activities in real time due to the latencies involved in enriching the current activities with the knowledge gained through batch processes. Therefore, the opportunity to provide tailored offers or showcase products based on customers’ preference and behaviors in near-real time, which contributes to a much better customer experience, was missing. Similarly, the opportunity to catch fraud within a second, before checkout, was also missing.
To improve the customer experience, Poshmark decided to invest in building a real-time analytics platform to enable real-time capabilities, as explained further in this post. Poshmark engineers worked closely with AWS architects through the AWS Data Lab program. The AWS Data Lab offers accelerated, joint engineering engagements between customers and AWS technical resources to create tangible deliverables that accelerate data and analytics modernization initiatives. The Design Lab is one half to two day engagement with customer team offering prescriptive guidance to arrive at the optimal solution architecture design before you embark on building the platform.
Designing the solution architecture through the AWS Data Lab process
The business and technical stakeholders from Poshmark and the AWS Data Lab architects discussed near-to-long-term business requirements along with the functional and non-functional capabilities required to decide on the architecture approach. They reviewed the current state architecture and constraints to understand data flow and technical integration points. The joint team discussed the pros and cons of various AWS services that already exist in Poshmark’s current architecture, as well as other AWS services that can meet the requirements.
Poshmark wanted to address the following business use cases via the real-time analytics platform:
Sessionization – Poshmark captures both server-side application events and client-side tracking events. They wanted to use these events to identify and analyze user sessions to track behavior.
Illegitimate sign-up and sign-in prevention – Poshmark wanted to detect and ban illegitimate sign-up or sign-in events from bots or non-human traffic in real time on the Poshmark application.
IP translation – The IP addresses present in events will be translated to city, state, and zip, and enriched with other information to implement near-real-time, location-aware services encompassing security-related functions as well as personalization functions.
Anonymization – Poshmark wanted to anonymize events and make the data available for internal users for querying in near-real time.
Personalized recommendations – User behavior based on clickstream events can be captured up to the last second before enriching it for personalization and sending it to the model to predict the recommendations.
Other use cases – Additional use cases relating to aggregations and machine learning (ML) inference use cases such as authorization to operate, listing spam detection, and avoiding account takeovers (ATOs), among others.
One common pattern identified for these use cases was the need for a central data enrichment pipeline to enrich incoming raw events before event data can be utilized for actual business processing. In the Design Lab, we focused on design for data enrichment pipelines aimed at enriching events with data from static files, dynamic data stores such as databases, APIs, or within the same event stream for the aforementioned streaming use cases. Later in this post, we cover the salient points discussed during the lab around design and architecture.
Batch analytics solution architecture
The following diagram shows the previous architecture at Poshmark. For brevity, only the flow pertaining to the real-time analytics platform is explained.
User interactions on Poshmark web and mobile applications generate server-side events. These events include add to cart, orders, transactions, and more on application servers, and the page view, clicks, and more on tracking servers. Fluentd with an Amazon Kinesis plugin is set up on both the application and tracking servers to send these events to Amazon Kinesis Data Streams. The Fluentd Kinesis plugin aggregates events before sending to Kinesis Data Streams. A single Kinesis data stream is currently set up to capture these events. A random partition key is configured in Fluentd for the events to allow even distribution of events across shards. The event data format is nested JSON. Poshmark maintains the same schema grammar at the first level of JSON for both server-side and client-side server events. The attributes at nested level can differ between server-side and client-side events.
Poshmark receives around 1 billion events per day (100 million per hour during peak hours, 10 million per hour during non-peak hours). The average size of the event record is 1.2 KB.
The data from the Kinesis data stream is consumed by two applications:
A Spark streaming application on Amazon EMR is used to write data from the Kinesis data stream to a data lake hosted on Amazon Simple Storage Service (Amazon S3) in a partitioned way. The data from the S3 data lake is used for batch processing and analytics through Amazon EMR and Amazon Redshift.
Druid hosted on Amazon Elastic Compute Cloud (Amazon EC2) integrates with the Kinesis data stream for streaming ingestion and allows users to run slice-and-dice OLAP queries. Operational dashboards are hosted on Grafana integrated with Druid.
Desired enhancements to the initial solution
The use cases discussed during the architecture sessions fall into one or more combinations of the following stream processing requirements:
Stateless event processing – For example, near-real-time anonymization.
External lookup – Looking up a value from external stores. For example, IP address, city, zip, state, or ID.
Stateful data processing – Accessing past events or aggregations or ML inferences.
To meet these requirements, the streaming platform is divided into two layers:
Central data enrichment – This layer runs enrichments commonly required by downstream streaming applications. This will help avoid replication of the same enrichment logic in each application and enable better operational maintenance. The enrichment should strive for per-record processing in most cases.
Specific streaming applications – This layer will house specific streaming applications with respect to use cases and utilize enriched data from the central data enrichment pipeline.
For central data enrichment, we made the following enhancements to the platform:
The total latency including ingestion and data enrichment was super critical and should be in the range of double-digit millisecond latency based on the overall latency budget of Poshmark to achieve real-time ML responses to events. The absolute lowest ingestion latency was achieved by Kafka, and the team decided to go with the managed version of Kafka, Amazon MSK.
Similarly, low-latency processing of data is also required, and appropriate framework should be considered accordingly.
Exactly-once delivery guarantees were required to avoid data duplication resulting in wrong calculations.
The enrichment source could be any source such as static files, databases, and APIs and latencies can vary between them. A number of server-side and client-side events are generated when a user interacts with a Poshmark application. As a result, the same information from the enrichment source is required to enrich each event. This frequently accessed information cached in a centralized cache will optimize fetch time.
Design decisions for the new solution
Poshmark made the following design decisions for central data enrichment:
Kafka can support double-digit millisecond latency from producer to consumer with appropriate performance tuning. Kafka can provide exactly-once semantics both at producers and consumer applications. AWS provides Kafka as part of its Amazon MSK offering, eliminating the operational overhead of maintaining and running Kafka cluster infrastructure on AWS, thereby allowing you to focus on developing and running Kafka-based applications. Poshmark decided to use Amazon MSK for their streaming ingestion and storage requirements.
We also decided to use Flink for streaming data enrichment applications for the following reasons:
Flink can provide low-latency processing even at higher throughput with exactly-once guarantees. Spark Structured Streaming on the other hand can provide low latency with low throughput due to microbatch-based processing. Spark Structured Streaming continuous processing is an experimental feature and provides at-least once guarantees.
The enrichment requests call to an external store if modeled in a map function (Spark’s map API or Flink’s MapFunction API) will make synchronous calls to the external store. The call will wait for a response from the external store before processing the next event, adding to delays and reducing overall throughput. The asynchronous interaction will allow sending requests and receiving responses concurrently from external stores. This will reduce wait time and improve overall throughput. Flink supports async I/O operators natively, allowing users to use asynchronous request clients with data streams. The API handles the integration with data streams, well as handling order, event time, fault tolerance, and more. Spark Structured Streaming doesn’t provide any such support natively and leaves it to users for custom implementation.
Poshmark selected Kinesis Data Analytics for Apache Flink to run the data enrichment application. Kinesis Data Analytics for Apache Flink provides the underlying infrastructure for your Apache Flink applications. It handles core capabilities like provisioning compute resources, parallel computation, automatic scaling, and application backups (implemented as checkpoints and snapshots).
An enrichment microservice accompanying Amazon ElastiCache for Redis was set up to abstract access from data enrichment applications. The AsyncFunction in the Flink async I/O operator isn’t multi-threaded and won’t work in a truly asynchronous way if the call is blocked or waiting for a response. The enrichment microservice handles requests and responses asynchronously coming from Flink async I/O operators. The data is also cached in ElastiCache for Redis to improve the latency of the microservice.
The Poshmark ML applications are the consumers of this enriched data. The team has built and deployed different ML models over time. These models include a learning to rank algorithm, fraud detection, personalization and recommendations, and online spam filtering. Previously, for deploying each model into production, the Poshmark team had to go through a series of infrastructure setup steps that involved data extraction from real-time sources, building real-time aggregate features from streaming data, storing these features in a low-latency database (Redis) for sub-millisecond inferences, and finally performing inferences via Amazon SageMaker hosted endpoints.
We also designed an ML feature storage pipeline that consumes data from the enriched streaming sources (Kinesis or Kafka), generate single-level and aggregated-level features, and ingest these generated features into a feature store repository with a very low latency of less than 80 milliseconds.
The ML models are now able to extract the needed features with latency less than 10 milliseconds from the feature repository and perform real-time model inferencing.
Real-time analytics solution architecture
The following diagram illustrates the solution architecture for real-time analytics with Amazon MSK and Kinesis Data Analytics for Apache Flink.
The workflow is as follows:
Users interact on Poshmark’s web or mobile application.
Server-side events are captured on application servers and client-side events are captured on tracking servers. These events are written in the downstream MSK cluster.
The raw events will be ingested into the MSK cluster using the Fluentd plugin to produce data for Kafka.
The enrichment microservice consists of reactive (asynchronous) enrichment lookup APIs fetching data from persistent data stores. ElastiCache for Redis caches frequently accessed data, reducing fetch time for enrichment lookup APIs.
The Flink application running on Kinesis Data Analytics for Apache Flink consumes raw events from Amazon MSK and runs data enrichment on a per-record basis. The Flink data enrichment application uses Flink’s async I/O to read external data from the enrichment lookup store for enriching stream events.
Enriched events are written in the MSK cluster under different enriched events topics.
The existing Spark streaming application consumes from the enriched events topic (or raw events topic) in Amazon MSK and writes the data into an S3 data lake.
Druid streaming ingestion now reads from the enriched events topic or raw events topic in Amazon MSK depending on the requirements.
Enrichment of the captured event data
In this section, we discuss the different steps to enrich the captured event data.
Kinesis Data Analytics for Apache Flink provides the underlying infrastructure for the Apache Flink applications. It handles core capabilities like provisioning compute resources, parallel computation, automatic scaling, and application backups (implemented as checkpoints and snapshots). You can use the high-level Flink programming features (such as operators, functions, sources, and sinks) in the same way that you use them when hosting the Flink infrastructure yourself.
Flink on Amazon EMR gives the flexibility to choose your Flink version, installation, configuration, instances, and storage. However, you also have to take care of cluster management and operational requirements such as scaling, application backup, and provisioning.
Enrichment lookup store
The AsyncFunction in the Flink async I/O operator isn’t multi-threaded and won’t work in a truly asynchronous way if the call is blocked or waiting for a response. The enrichment lookup API should handle requests and responses asynchronously coming from Flink async I/O operators. The enrichment lookup API can be hosted on Amazon EC2 or containers such as Amazon Elastic Container Service (Amazon ECS) or Amazon Elastic Kubernetes Service (Amazon EKS).
A number of server-side and client-side events are generated when a user interacts with a Poshmark application. As a result, the same information is required to enrich each event. This frequently accessed information cached in a centralized cache can optimize fetch time. The latency to the centralized cache can be further reduced by hosting the client (enrichment lookup API) and cache server in the same Availability Zone.
Reconciliation in case of pipeline errors
The event enrichment can fail in data enrichment applications for various reasons, such as the external store timing out or missing information in the store. The enriched fields may or may not be critical for downstream streaming applications. You should build your downstream streaming applications considering that these failures can occur and implement a fallback mechanism, for example retrying on-demand enrichment from the application. The failure handling will also be governed by latency tolerance of the application.
The processing of data is based on event time. In some situations, data can arrive late in the platform. Both Flink and Spark allow lateness and watermarks for users to handle late-arriving data by defining thresholds. Late-arriving data beyond the threshold is discarded from processing. It’s possible to get this discarded too-late data in Flink using a side output. There is no such provision in Spark Structured Streaming.
A few streaming applications require their batch counterpart to reconcile data hourly or daily to handle data mismatch or data discrepancy due to late-arriving data or missing data.
Improved customer experience
The new real-time architecture offered the following benefits for an improved customer experience:
Anonymization – Poshmark is now able to provide and utilize real-time anonymized data for multiple functions both internally and externally because anonymization happens in real time.
Fraud mitigation – Poshmark was previously able to detect and prevent 45% of ATOs with the batch-based solution. With the real-time system, Poshmark is able to prevent 80% of ATOs.
Personalization – By providing personalized search results, Poshmark achieved an 8% improvement on clickthrough rates for search. This is a significant increase in the top of the funnel, increasing overall search conversions.
Improvement in these three factors helped end-customers gain confidence in the Poshmark app and website, which in turn enabled customers to increase their interaction with the app and helped accelerate customer engagement and growth.
In this post, we discussed the ingestion of real-time clickstream and log event data into Amazon MSK. We showed how enrichment of the captured data can be performed through Kinesis Data Analytics for Apache Flink. We broke up the enrichment processing into multiple components, such as Kinesis Data Analytics for Apache Flink, the enrichment microservices and the enrichment lookup store, and an enrichment cache. We discussed the downstream applications that used this enriched customer information to perform real-time security checks and offer personalized recommendations to end-users. We also discussed some of the areas that may need attention in case there are failures in the pipeline. Lastly, we showed how Poshmark improved their customer experience and gained market share by implementing this real-time analytics pipeline.
About the authors
Mahesh Pasupuleti is a VP of Data & Machine Learning Engineering at Poshmark. He has helped several startups succeed in different domains, including media streaming, healthcare, the financial sector, and marketplaces. He loves software engineering, building high performance teams, and strategy, and enjoys gardening and playing badminton in his free time.
Gaurav Shah is Director of Data Engineering and ML at Poshmark. He and his team help build data-driven solutions to drive growth at Poshmark.
Raghu Mannam is a Sr. Solutions Architect at AWS in San Francisco. He works closely with late-stage startups, many of which have had recent IPOs. His focus is end-to-end solutioning including security, DevOps automation, resilience, analytics, machine learning, and workload optimization in the cloud.
Deepesh Malviya is Solutions Architect Manager on the AWS Data Lab team. He and his team help customers architect and build data, analytics, and machine learning solutions to accelerate their key initiatives as part of the AWS Data Lab.