Case study · Stack

How PicPay scaled a payments super app on AWS

PicPay started as a peer-to-peer payments wallet and grew into a full digital bank. It was built on AWS from early on, with apps on Kubernetes in EKS, Kafka on MSK, and a data platform that moved to Databricks.
Copy this stackpicpay.com Digital wallet and bank · Brazil · founded 2012

Hosting

  • AWSbaguete.com.br

    Cloud that ran 100% of PicPay's operation as of 2018.

    Why: Being fully on AWS let PicPay make the most of explosive growth while keeping the experience good.

  • Kubernetes (Amazon EKS)sreteams.substack.com

    Runs the applications.

    Why: The SRE team prefers AWS services that perform well with little management.

Database

Build and runtime

  • Harnesssreteams.substack.com

    Deploys new Docker images to the EKS clusters after CodeBuild.

  • Gradle Develocitydevelocity.ai

    Build cache and build scans for the Android builds, locally and in CI.

    Why: Developers were losing too much time waiting on Android build and test feedback.

Data fetching

  • Apache Kafka (Amazon MSK)sreteams.substack.com

    Event streaming, also used to carry logs.

  • Amazon ElastiCachesreteams.substack.com

    Managed cache for the applications.

  • Databricksdatabricks.com

    Lakehouse that unifies transactions, sales, finance and logs for BI and ML.

    Why: The old multi-tool stack on AWS was too complex, slow and hard to share.

Backend framework

AI model

  • Rasarasa.com

    Conversational assistant that guided users through COVID-19 emergency aid.

    Why: It beat IBM Watson on cost and accuracy, is open source and handled Portuguese well.

Analytics

  • Tableaudatabricks.com

    Dashboards on top of the Databricks data.

  • Mixpanelmixpanel.com

    Tracks the signup funnel, A/B tests and individual user journeys.

    Why: A new product manager can learn it fast, without writing queries or building dashboards by hand.

Error monitoring

  • New Relicnewrelic.com

    Centralizes logs, telemetry, alerts and executive dashboards.

    Why: Incident handling had become slow and the team was buried in alerts.

  • OpsGeniesreteams.substack.com

    Pages the on-call engineer from monitoring alerts.

WhatsApp and SMS

  • Sendbirdsendbird.com

    Powers in-app chat around payments.

    Why: PicPay wanted engagement beyond single tasks, and social users transact twice as much.

How they built it

All in on AWS from early on

By 2018 PicPay ran entirely on AWS and used more than 20 of its services, including API Gateway, DynamoDB, ElastiCache, Lambda, RDS and VPC. The COO credited that setup with letting the company ride 2,000% growth in five months in 2017 without hurting the user experience.

Sources:baguete.com.br

A platform built on managed services

PicPay's SRE team favors AWS services that perform well and need little management. In 2022 applications ran on Kubernetes in EKS with some Lambda functions, data sat in RDS, Kafka ran on MSK, and SQS, SNS and ElastiCache covered queues and cache.

CodeBuild runs tests and pushes Docker images to ECR, and Harness rolls them out to the EKS clusters. Filebeat and Logstash stream logs into Kafka, which forwards them to the logging tool and backs them up to S3 Glacier, and OpsGenie pages whoever is on call.

Sources:sreteams.substack.com

Fewer incidents with New Relic

As PicPay grew, incident handling got slow and the team suffered from alert fatigue. It standardized alerts and pulled logs and telemetry into New Relic, which cut incidents by 65%, improved time to resolution by 30% and reduced yearly downtime by 51%.

Sources:newrelic.com

From many data tools to one lakehouse

The first data stack on AWS stitched many tools together. Pipelines were slow, teams could not share data easily and analysts had no real-time view. PicPay moved to Databricks with Delta Lake and Databricks SQL, which now feeds more than 50 dashboards, cash-back targeting models and fraud monitoring.

PicPay reports $10M saved through smarter cash-back and 50% lower analytics and visualization costs.

Sources:databricks.com

Tools inside the product

Chat built on Sendbird turns payments into conversations, since socially active users transact twice as much. Mixpanel lets the growth team watch the eight-step signup funnel hour by hour, and a clearer password message found that way lifted conversion by 10%.

During Brazil's COVID-19 emergency aid, PicPay shipped a Rasa assistant in about ten business days to guide users through eligibility. On Android, Gradle's Develocity build cache and build scans sped up slow build and test cycles.

Sources:sendbird.com mixpanel.com rasa.com develocity.ai

Also detected on their website

Found on the homepage with the open Wappalyzer rules (HTML, scripts and headers). It shows what runs in the browser, not the whole backend.

  • Amazon CloudFront
  • Amazon Web Services
  • HSTS
  • HTTP/3
  • Varnish