Alternatives to Druid

Recently we were looking at a few robust and cost-effective ways of replicating the data that resides in our production MongoDB to a PostgreSQL database for data warehousing and business intelligence.

We set ourselves the following criteria for the optimal tool that would do this job: - The data replication must be near real-time, yet it should NOT impact the production database - The data replication must be horizontally scalable (based on the load), asynchronous & crash-resilient

Based on the above criteria, we selected the following tools to perform the end to end data replication:

We chose MongoDB Stitch for picking up the changes in the source database. It is the serverless platform from MongoDB. One of the services offered by MongoDB Stitch is Stitch Triggers. Using stitch triggers, you can execute a serverless function (in Node.js) in real time in response to changes in the database. When there are a lot of database changes, Stitch automatically "feeds forward" these changes through an asynchronous queue.

We chose Amazon SQS as the pipe / message backbone for communicating the changes from MongoDB to our own replication service. Interestingly enough, MongoDB stitch offers integration with AWS services.

In the Node.js function, we wrote minimal functionality to communicate the database changes (insert / update / delete / replace) to Amazon SQS.

Next we wrote a minimal micro-service in Python to listen to the message events on SQS, pickup the data payload & mirror the DB changes on to the target Data warehouse. We implemented source data to target data translation by modelling target table structures through SQLAlchemy . We deployed this micro-service as AWS Lambda with Zappa. With Zappa, deploying your services as event-driven & horizontally scalable Lambda service is dumb-easy.

In the end, we got to implement a highly scalable near realtime Change Data Replication service that "works" and deployed to production in a matter of few days!

We use MongoDB as our primary #datastore. Mongo's approach to replica sets enables some fantastic patterns for operations like maintenance, backups, and #ETL.

As we pull #microservices from our #monolith, we are taking the opportunity to build them with their own datastores using PostgreSQL. We also use Redis to cache data we’d never store permanently, and to rate-limit our requests to partners’ APIs (like GitHub).

When we’re dealing with large blobs of immutable data (logs, artifacts, and test results), we store them in Amazon S3. We handle any side-effects of S3’s eventual consistency model within our own code. This ensures that we deal with user requests correctly while writes are in process.

1.0 of Stream leveraged Cassandra for storing the feed. Cassandra is a common choice for building feeds. Instagram, for instance started, out with Redis but eventually switched to Cassandra to handle their rapid usage growth. Cassandra can handle write heavy workloads very efficiently.

Cassandra is a great tool that allows you to scale write capacity simply by adding more nodes, though it is also very complex. This complexity made it hard to diagnose performance fluctuations. Even though we had years of experience with running Cassandra, it still felt like a bit of a black box. When building Stream 2.0 we decided to go for a different approach and build Keevo. Keevo is our in-house key-value store built upon RocksDB, gRPC and Raft.

RocksDB is a highly performant embeddable database library developed and maintained by Facebook’s data engineering team. RocksDB started as a fork of Google’s LevelDB that introduced several performance improvements for SSD. Nowadays RocksDB is a project on its own and is under active development. It is written in C++ and it’s fast. Have a look at how this benchmark handles 7 million QPS. In terms of technology it’s much more simple than Cassandra.

This translates into reduced maintenance overhead, improved performance and, most importantly, more consistent performance. It’s interesting to note that LinkedIn also uses RocksDB for their feed.

#InMemoryDatabases #DataStores #Databases

React AngularJS jQuery

Laravel Zend Framework

MySQL MongoDB Cassandra

Docker

Linux

Why we spent several years building an open source, large-scale metrics alerting system, M3, built for Prometheus:

By late 2014, all services, infrastructure, and servers at Uber emitted metrics to a Graphite stack that stored them using the Whisper file format in a sharded Carbon cluster. We used Grafana for dashboarding and Nagios for alerting, issuing Graphite threshold checks via source-controlled scripts. While this worked for a while, expanding the Carbon cluster required a manual resharding process and, due to lack of replication, any single node’s disk failure caused permanent loss of its associated metrics. In short, this solution was not able to meet our needs as the company continued to grow.

To ensure the scalability of Uber’s metrics backend, we decided to build out a system that provided fault tolerant metrics ingestion, storage, and querying as a managed platform...

https://eng.uber.com/m3/

(GitHub : https://github.com/m3db/m3)

We have Prometheus as a monitoring engine as a part of our stack which contains Kubernetes cluster, container images and other open source tools. Also, I am aware that Sysdig can be integrated with Prometheus but I really wanted to know whether Sysdig or sysdig+prometheus will make better monitoring solution.

Back in 2014, I was given an opportunity to re-architect SmartZip Analytics platform, and flagship product: SmartTargeting. This is a SaaS software helping real estate professionals keeping up with their prospects and leads in a given neighborhood/territory, finding out (thanks to predictive analytics) who's the most likely to list/sell their home, and running cross-channel marketing automation against them: direct mail, online ads, email... The company also does provide Data APIs to Enterprise customers.

I had inherited years and years of technical debt and I knew things had to change radically. The first enabler to this was to make use of the cloud and go with AWS, so we would stop re-inventing the wheel, and build around managed/scalable services.

For the SaaS product, we kept on working with Rails as this was what my team had the most knowledge in. We've however broken up the monolith and decoupled the front-end application from the backend thanks to the use of Rails API so we'd get independently scalable micro-services from now on.

Our various applications could now be deployed using AWS Elastic Beanstalk so we wouldn't waste any more efforts writing time-consuming Capistrano deployment scripts for instance. Combined with Docker so our application would run within its own container, independently from the underlying host configuration.

Storage-wise, we went with Amazon S3 and ditched any pre-existing local or network storage people used to deal with in our legacy systems. On the database side: Amazon RDS / MySQL initially. Ultimately migrated to Amazon RDS for Aurora / MySQL when it got released. Once again, here you need a managed service your cloud provider handles for you.

Future improvements / technology decisions included:

Caching: Amazon ElastiCache / Memcached CDN: Amazon CloudFront Systems Integration: Segment / Zapier Data-warehousing: Amazon Redshift BI: Amazon Quicksight / Superset Search: Elasticsearch / Amazon Elasticsearch Service / Algolia Monitoring: New Relic

As our usage grows, patterns changed, and/or our business needs evolved, my role as Engineering Manager then Director of Engineering was also to ensure my team kept on learning and innovating, while delivering on business value.

One of these innovations was to get ourselves into Serverless : Adopting AWS Lambda was a big step forward. At the time, only available for Node.js (Not Ruby ) but a great way to handle cost efficiency, unpredictable traffic, sudden bursts of traffic... Ultimately you want the whole chain of services involved in a call to be serverless, and that's when we've started leveraging Amazon DynamoDB on these projects so they'd be fully scalable.

Although we were using Elasticsearch in the beginning to power our in-app search, we moved this part of our processing over to Algolia a couple of months ago; this has proven to be a fantastic choice, letting us build search-related features with more confidence and speed.

Elasticsearch is only used for searching in internal tooling nowadays; hosting and running it reliably has been a task that took up too much time for us in the past and fine-tuning the results to reach a great user-experience was also never an easy task for us. With Algolia we can flexibly change ranking methods on the fly and can instead focus our time on fine-tuning the experience within our app.

Memcached is used in front of most of the API endpoints to cache responses in order to speed up response times and reduce server-costs on our side.

#SearchAsAService

The algorithms and data infrastructure at Stitch Fix is housed in #AWS. Data acquisition is split between events flowing through Kafka, and periodic snapshots of PostgreSQL DBs. We store data in an Amazon S3 based data warehouse. Apache Spark on Yarn is our tool of choice for data movement and #ETL. Because our storage layer (s3) is decoupled from our processing layer, we are able to scale our compute environment very elastically. We have several semi-permanent, autoscaling Yarn clusters running to serve our data processing needs. While the bulk of our compute infrastructure is dedicated to algorithmic processing, we also implemented Presto for adhoc queries and dashboards.

Beyond data movement and ETL, most #ML centric jobs (e.g. model training and execution) run in a similarly elastic environment as containers running Python and R code on Amazon EC2 Container Service clusters. The execution of batch jobs on top of ECS is managed by Flotilla, a service we built in house and open sourced (see https://github.com/stitchfix/flotilla-os).

At Stitch Fix, algorithmic integrations are pervasive across the business. We have dozens of data products actively integrated systems. That requires serving layer that is robust, agile, flexible, and allows for self-service. Models produced on Flotilla are packaged for deployment in production using Khan, another framework we've developed internally. Khan provides our data scientists the ability to quickly productionize those models they've developed with open source frameworks in Python 3 (e.g. PyTorch, sklearn), by automatically packaging them as Docker containers and deploying to Amazon ECS. This provides our data scientist a one-click method of getting from their algorithms to production. We then integrate those deployments into a service mesh, which allows us to A/B test various implementations in our product.

For more info:

• Our Algorithms Tour: https://algorithms-tour.stitchfix.com/
• Our blog: https://multithreaded.stitchfix.com/blog/
• Careers: https://multithreaded.stitchfix.com/careers/

#DataScience #DataStack #Data

Why we built Marmaray, an open source generic data ingestion and dispersal framework and library for Apache Hadoop :

Built and designed by our Hadoop Platform team, Marmaray is a plug-in-based framework built on top of the Hadoop ecosystem. Users can add support to ingest data from any source and disperse to any sink leveraging the use of Apache Spark . The name, Marmaray, comes from a tunnel in Turkey connecting Europe and Asia. Similarly, we envisioned Marmaray within Uber as a pipeline connecting data from any source to any sink depending on customer preference:

https://eng.uber.com/marmaray-hadoop-ingestion-open-source/

(Direct GitHub repo: https://github.com/uber/marmaray Kafka Kafka Manager )

I use Amazon Athena because similar to Google BigQuery , you can store and query data easily. Especially since you can define data schema in the Glue data catalog, there's a central way to define data models.

However, I would not recommend for batch jobs. I typically use this to check intermediary datasets in data engineering workloads. It's good for getting a look and feel of the data along its ETL journey.

I need to build the Alert & Notification framework with the use of a scheduled program. We will analyze the events from the database table and filter events that are falling under a day timespan and send these event messages over email. Currently, we are using Kafka Pub/Sub for messaging. The customer wants us to move on Apache Flink, I am trying to understand how Apache Flink could be fit better for us.

To provide employees with the critical need of interactive querying, we’ve worked with Presto, an open-source distributed SQL query engine, over the years. Operating Presto at Pinterest’s scale has involved resolving quite a few challenges like, supporting deeply nested and huge thrift schemas, slow/ bad worker detection and remediation, auto-scaling cluster, graceful cluster shutdown and impersonation support for ldap authenticator.

Our infrastructure is built on top of Amazon EC2 and we leverage Amazon S3 for storing our data. This separates compute and storage layers, and allows multiple compute clusters to share the S3 data.

We have hundreds of petabytes of data and tens of thousands of Apache Hive tables. Our Presto clusters are comprised of a fleet of 450 r4.8xl EC2 instances. Presto clusters together have over 100 TBs of memory and 14K vcpu cores. Within Pinterest, we have close to more than 1,000 monthly active users (out of total 1,600+ Pinterest employees) using Presto, who run about 400K queries on these clusters per month.

Each query submitted to Presto cluster is logged to a Kafka topic via Singer. Singer is a logging agent built at Pinterest and we talked about it in a previous post. Each query is logged when it is submitted and when it finishes. When a Presto cluster crashes, we will have query submitted events without corresponding query finished events. These events enable us to capture the effect of cluster crashes over time.

Each Presto cluster at Pinterest has workers on a mix of dedicated AWS EC2 instances and Kubernetes pods. Kubernetes platform provides us with the capability to add and remove workers from a Presto cluster very quickly. The best-case latency on bringing up a new worker on Kubernetes is less than a minute. However, when the Kubernetes cluster itself is out of resources and needs to scale up, it can take up to ten minutes. Some other advantages of deploying on Kubernetes platform is that our Presto deployment becomes agnostic of cloud vendor, instance types, OS, etc.

#BigData #AWS #DataScience #DataEngineering

The algorithms and data infrastructure at Stitch Fix is housed in #AWS. Data acquisition is split between events flowing through Kafka, and periodic snapshots of PostgreSQL DBs. We store data in an Amazon S3 based data warehouse. Apache Spark on Yarn is our tool of choice for data movement and #ETL. Because our storage layer (s3) is decoupled from our processing layer, we are able to scale our compute environment very elastically. We have several semi-permanent, autoscaling Yarn clusters running to serve our data processing needs. While the bulk of our compute infrastructure is dedicated to algorithmic processing, we also implemented Presto for adhoc queries and dashboards.

Beyond data movement and ETL, most #ML centric jobs (e.g. model training and execution) run in a similarly elastic environment as containers running Python and R code on Amazon EC2 Container Service clusters. The execution of batch jobs on top of ECS is managed by Flotilla, a service we built in house and open sourced (see https://github.com/stitchfix/flotilla-os).

At Stitch Fix, algorithmic integrations are pervasive across the business. We have dozens of data products actively integrated systems. That requires serving layer that is robust, agile, flexible, and allows for self-service. Models produced on Flotilla are packaged for deployment in production using Khan, another framework we've developed internally. Khan provides our data scientists the ability to quickly productionize those models they've developed with open source frameworks in Python 3 (e.g. PyTorch, sklearn), by automatically packaging them as Docker containers and deploying to Amazon ECS. This provides our data scientist a one-click method of getting from their algorithms to production. We then integrate those deployments into a service mesh, which allows us to A/B test various implementations in our product.

For more info:

• Our Algorithms Tour: https://algorithms-tour.stitchfix.com/
• Our blog: https://multithreaded.stitchfix.com/blog/
• Careers: https://multithreaded.stitchfix.com/careers/

#DataScience #DataStack #Data

To provide employees with the critical need of interactive querying, we’ve worked with Presto, an open-source distributed SQL query engine, over the years. Operating Presto at Pinterest’s scale has involved resolving quite a few challenges like, supporting deeply nested and huge thrift schemas, slow/ bad worker detection and remediation, auto-scaling cluster, graceful cluster shutdown and impersonation support for ldap authenticator.

Our infrastructure is built on top of Amazon EC2 and we leverage Amazon S3 for storing our data. This separates compute and storage layers, and allows multiple compute clusters to share the S3 data.

We have hundreds of petabytes of data and tens of thousands of Apache Hive tables. Our Presto clusters are comprised of a fleet of 450 r4.8xl EC2 instances. Presto clusters together have over 100 TBs of memory and 14K vcpu cores. Within Pinterest, we have close to more than 1,000 monthly active users (out of total 1,600+ Pinterest employees) using Presto, who run about 400K queries on these clusters per month.

Each query submitted to Presto cluster is logged to a Kafka topic via Singer. Singer is a logging agent built at Pinterest and we talked about it in a previous post. Each query is logged when it is submitted and when it finishes. When a Presto cluster crashes, we will have query submitted events without corresponding query finished events. These events enable us to capture the effect of cluster crashes over time.

Each Presto cluster at Pinterest has workers on a mix of dedicated AWS EC2 instances and Kubernetes pods. Kubernetes platform provides us with the capability to add and remove workers from a Presto cluster very quickly. The best-case latency on bringing up a new worker on Kubernetes is less than a minute. However, when the Kubernetes cluster itself is out of resources and needs to scale up, it can take up to ten minutes. Some other advantages of deploying on Kubernetes platform is that our Presto deployment becomes agnostic of cloud vendor, instance types, OS, etc.

#BigData #AWS #DataScience #DataEngineering