Apache Kafka® is at the heart of the data transportation layer at Pinterest. The amount of data that runs through Kafka has constantly grown over the years. This growth sometimes brings operational challenges that we have to deal with and plan for to make sure data transportation runs as smoothly as possible. This article shares how we run Kafka at Pinterest and discusses some of the challenges we’ve faced and how we’ve addressed them.
Pinterest is a visual discovery engine for finding ideas like recipes, home and style inspiration, and more. The user base and the use cases that attract users to the platform have dramatically increased over the years. As of December 2020, Pinterest attracts over 459 million monthly active users and hosts 300 billion pins and 6.7+ billion boards—and these numbers continue to grow. This growth has resulted in an ever-increasing amount of generated data that requires transportation from the source to various systems, real time and batch, for ingestion and processing.
The diagram above illustrates how data moves at Pinterest. Events associated with user actions on the platform (mobile application or web) are logged in files. Singer, our highly performant and efficient logging agent, runs agents on hosts that are responsible for sending those logs to Kafka. The other path that flows messages to Kafka is the incremental database ingestion pipeline that is powered by Maxwell, which is responsible for sending database changelogs to Kafka for downstream use cases.
There are two types of use cases for messages in Kafka clusters:
The Logging Platform team manages the topology components, from Singer to S3 Transporter and Sanitizer, and provides the necessary interfaces to its clients to make use of the service. More on this is covered in the automation section below. The diagram also shows the critical role that Kafka plays in this topology: a centerpiece that glues all of the other components together.
The Logging Platform runs and manages Kafka on Amazon EC2 virtual machines across multiple regions. The number of production clusters is over 50 and continues to grow over time. Clusters are classified based on domains, use case families, and impact radius.
As of December 2020, these 50+ clusters host 3,000+ brokers, 3,000+ topics, and ~500K partitions (including replicas). The inbound and outbound traffic peaks at ~25 GB/s and 50 GB/s, respectively (inbound messages peak at over 40 million/s). All clusters run Kafka version 2.3.1 with some cherry-picked commits.
As Pinterest grows and onboards more and more users, so does the amount of data that needs to be transported from producers or logs to downstream services. This means that our services need to handle organic growth as well as use case growth, which has presented several challenges for the Logging Platform over time. This section briefly goes over these challenges and explains how we addressed them.
The hardware that a virtual machine (VM) runs on does not last forever and eventually has to be recycled. The Kafka broker that runs on the VM also goes out of service and needs to start fresh on another VM, replicating all the data it had before the host went down. Given the number of brokers that we manage, each week, a few brokers will go through the recovery stage. There are also maintenance operations, such as broker upgrades, OS upgrades, and rolling restarts, which require partial or full replays of logs for in-sync replica (ISR) restoration.
We used to run Kafka brokers on D2 instance types with magnetic disks. We also used to perform partition reassignments by moving partitions out of the broker that was going to be replaced and waiting for its full recovery. In this situation, the replaced broker has to perform disk I/O operations at a much-higher-than-usual rate. Because classic magnetic disks have limited IOPS during high load times (specifically during recovery), the queue depth for IOPS increases, causing increased latency, which triggers a cascading effect across topics and an entire cluster. For example, if the replaced broker, b1, has to fetch several partitions from the same leader broker, b2, it impacts the network bandwidth of b2, causing it to drop from the ISRs of the partitions that it is following. This chain reaction quickly brings several partitions down to a single ISR.
We looked for ways to improve the broker recovery process and reduce its impact on the whole cluster. After performing a root cause analysis of this issue, we discovered that the CPU I/O wait times of brokers spiked during these degraded states. Upon further investigation, we were able to attribute this observation to the density of Kafka brokers, that is, the number of partitions on the broker, the load on each partition, and the spike in fetches during broker recovery.
The answer was clear: Magnetic drives were simply not good enough to sustain our environment and did not meet our latency and operations requirements. We switched to SSDs because they provide a few orders of magnitude higher IOPS, giving brokers the burst capacity needed for recovery and load spikes and allowing us to operate at even higher broker densities. This has brought our CPU I/O wait time down from 10–30% (during degraded state) to < 0.1% (p100), resolving the degraded state issue and providing nearly uninterrupted service even while under-replicated partitions (URPs) are being recovered.
In addition to fixing our operations issues, SSDs provide consistent and predictable performance to our customers, irrespective of the fan-out ratios (number of consumers) and spikes in traffic (organic or otherwise).
As mentioned in the previous section, we used to perform automated partition reassignments before each unhealthy broker replacement. This meant moving all partitions out of the unhealthy broker and temporarily assigning them to other brokers in the cluster and replacing the unhealthy node. Over time and after several incidents, we realized that the guarantee that this partition movement provides (a higher degree of availability for the partitions on the unhealthy broker for an average of about 30 minutes) was equal to the organic recovery of a Kafka broker.
As a result, we decided to disable the partition movement and let the unhealthy broker naturally recover after it is replaced. Instead, we transitioned to a static assignment model called brokersets. We’ve run this operation model for almost two years and have never regretted it.
Our Kafka clusters used to run on different default log message format versions. Since most topics do not override these versions, significant tech debt organically piled up over time. This is primarily because upgrading the default log message format version of a cluster meant making sure that no existing (potentially old) client would break due to not being able to communicate with the brokers. This resulted in having clusters with an old default log message format version and a variety of clients (old and new) that used those clusters. Given that we had already invested in building a lineage system, it was simpler for us to track clients and pursue the system-wide upgrade initiative.
In general, whenever the format version of messages sent to Kafka or expected to be received from Kafka is different from the format version of messages stored by the broker, the broker may potentially have to convert those messages to match the client’s expectations. This conversion means additional CPU cycles on the broker side, which is amplified if the batch of messages is compressed because the CPU cost of decompression and recompression is added to the CPU cost of message format conversion.
As we explain further in the Kafka upgrade section below, having old clients upgrade to a more recent Kafka client library and upgrading the default log message format version of all clusters help us reduce unnecessary CPU load from the brokers and allow us to size Kafka clusters more efficiently and reduce costs.
This section discusses some of the actions we took to reduce the overall cost of Kafka clusters. For a Kafka cluster, the costs include:
With Kafka clusters and their clients hosted on EC2 instances, data transfer costs associated with the clusters include the following:
The majority of these costs occur when data is transferred in or out of an AWS region’s availability zone (AZ). For availability reasons, we keep ISRs in separate AZs; therefore, the cost associated with replicating data between brokers in each partition’s ISR set is the price that we pay for ensuring higher availability. For producers and consumers, however, we’ve implemented a rack-aware partitioner on the producer side (for Singer) and a rack-aware partition assignment strategy on the consumer side (for S3 Transporter) to alleviate a large percentage of the data transfer cost for moving data in and out of Kafka clusters. For more on this, please refer to this blog post: Optimizing Kafka for the Cloud.
Sending and storing uncompressed data in Kafka clusters leads to an increase in two types of AWS costs:
By enforcing compression on the producer side, we lowered our infrastructure costs considerably on these fronts. This cost reduction outweighs the small increase of CPU load (< 10%) on the client side for having to deal with compressed messages.
We also use
SinglePartitionPartitioner (the producer only writes to a single randomly selected partition) to improve compression ratios by up to two times since larger batches are leveraged.
To further reduce the cost of Kafka clusters, we’ve implemented a few other improvements:
Having to deal with scattered and randomly distributed partitions of each topic within the cluster used to be a recurring challenge for us (we used the default placement strategy provided by the new topic creation API). If there was an issue with a particular topic (e.g., the traffic suddenly spiked on the topic without prior notice), the issue could easily cascade to all other topics and the cluster as a whole because partition placements were not isolated for different topics.
To limit the blast radius in situations like this, we decided to enforce static partition assignments for each topic. We created the concept of brokersets, a logical subset of brokers in a Kafka cluster to which one or more topics can be assigned for placement. If a topic is assigned to a brokerset, the partitions of that topic will only be placed on the brokers contained in that set. For example, if a brokerset contains brokers 1 to 6 over three AZs (AZ1:1, 4; AZ2: 2, 5; AZ3: 3, 6), and a six-partition topic with replication factor of 3 assigned to it, all of those partitions (all replicas) will be placed on brokers 1 to 6.
We also defined the notion of strides, which determines how the actual partition assignment is calculated within a brokerset. They factor in AZ placement as well to make sure partition leaders are balanced across AZs. A stride of 0 means sequential placement, while higher strides add gaps in the sequence of replicas. For example, for the above brokerset 1–6 and the six-partition topic assigned to it, a stride of 0 and 1 lead to these partition assignments:
In both cases, each broker is the leader of one partition and a follower of two partitions. Each AZ is the leader of two partitions and follower of four partitions. Both strides 0 and 1 provide a balanced distribution.
The significance of strides comes into play when multiple topics are placed on the same brokerset. If all topics on the above brokerset use stride = 0, and, for example, broker 1 goes under maintenance or replacement, all leaders that were on 1 will temporarily move to broker 2. This doubles the load on broker 2 and could very well degrade that broker by saturating its resources (network bandwidth, CPU load, etc.). If topics use different stride values and one broker goes down, its load will be distributed among the other brokers in a balanced fashion (this actually depends on both the brokerset size and partition count), which means a single broker no longer has to carry all that extra load. Our partition assignment algorithm easily calculates the assignments based on these factors.
Brokersets simplify topic scaling too. If a topic’s throughput is expected to scale up or down and the existing brokerset is no longer a good fit, the topic can be easily assigned to another brokerset that is ready to handle the expected load. This leads to a one-time partition reassignment but merely changes the isolation of the topic from one brokerset to another.
If we ignore the Kafka cluster controller, brokersets provide virtual Kafka clusters within a Kafka cluster. Note that brokersets do not have to cover a single sequential list of broker IDs; multiple broker sequences can be used to build a brokerset, providing us the ability to minimize data movement during topic expansions. We also use brokersets to perform incremental capacity allocation on the cluster: we create a new brokerset when all the existing ones are completely utilized and also provide this awareness in our topic onboarding wizards to be load aware so that we can minimize operational churn for our team.
Brokerset-based topic management is available as a plugin in Orion (see the automation section for details).
We created a unified management layer called Orion to manage Kafka and other stateful distributed systems like HBase.
Until a year ago, both the Logging Platform team and our customers used Yahoo’s CMAK (formerly known as Kafka Manager) to get a view of the Kafka clusters, individual topics, and their configuration. We also used to rely on our earlier (now deprecated) open sourced auto-healing solution, DoctorK, for auto-recovery of brokers. There were inherent problems with these two systems:
We implemented Orion as a generic solution for managing various stateful systems managed by our team. Orion, which we recently open sourced, is a unified solution that provides:
Pinterest uses Orion to manage all of its Kafka clusters and benefits from the following features:
The image below highlights screenshots of Orion in action. You can check out Orion on GitHub for more information.
Up until last year, we used CMAK and made it available to customers for creating and configuring their own topics. Unfortunately, this topic creation model led to several problems over time:
These issues were too much to ignore considering the fast-growing number of use cases and Kafka topics that had to be created to support them. In response, we moved toward a managed topic model to factor in the current state of the cluster before adding a new topic and also avoided additional overhead for customers to create or reconfigure topics.
We created an extension in our UI (Aerial) by adding a wizard for topic creation that asks for some critical information about the topic that is being created. It generates a PR that is sent to the Logging Platform for approval. Once approved and landed, Orion automatically creates the topic on the specified brokerset and notifies the customer of the topic creation. In most use cases, the number of partitions is flexible and the wizard comes up with the proper partition count and placement based on the projected topic throughout and the current cluster metrics and brokersets. Once created, Orion monitors the topic and its health and performs automated healing actions when necessary.
Once a pipeline is set up and traffic starts to flow in and out of a topic, customers can see a bird’s-eye view of that pipeline and the data flow among different components. The image below shows which Singer host the pipeline initiates from, which Kafka topic those logs are written to, and which S3 Transporter pipelines receive those logs. For each node in this data ingestion route, there are links that can point the customer to the related component’s configuration. Aerial tracks the end-to-end lineage for our team’s services and makes it available to our customers for self-service troubleshooting and investigation.
In 2020, we upgraded all our Kafka clusters to bring consistency to their broker version, their inter-broker protocol version, and their default log message format version. To review, here are the key pieces that were involved:
The following table color codes different versions of Kafka brokers, the inter-broker protocol, and log message format. It shows where we were before the upgrade (columns with a variety of colors) and where we are now (uniformly colored columns on the right) with our Kafka clusters. It shows that we were almost consistent prior to the upgrade with respect to the broker version in use (2.0.0). Ideally, we should have been similarly consistent on inter-broker protocol versions, as they should be upgraded right after broker upgrades, and there is no particular reason why broker versions are upgraded but inter-broker protocol versions are not. This was tech debt on our side that we resolved during the 2020 upgrades.
The main inconsistency, however, was with log message protocol versions, which was the riskiest part of the upgrade. Upgrading the log message format version required ensuring that there is no old client that could stop working after the upgrade or impose additional load on the cluster (i.e., due to log message format conversion). As mentioned earlier, these conversions pose additional CPU usage on the cluster, because the brokers have to decode and re-encode all the messages in a batch and could impact broker performance (latencies/throughput). This is particularly expensive for compressed topics. Upgrading old log message formats was the most time-consuming part of our upgrade since it involved digging into what Kafka client library version was using each topic. Note that KIP-511, which is fully included in 2.5.0, fills this gap from the Kafka broker point of view.
Based on our experience performing these upgrades, we’ve gathered some useful lessons to share.
Most major release versions that are filled with new features are not properly battle-tested before the release, and several bug fixes typically follow in minor releases. Therefore, it is safer to go with a bug fix release, which is why we chose 2.3.1 and not 2.3.0.
However, there is no guarantee that bugs won’t creep into bug fix releases. We actually ran into KAFKA-9752, which is a side effect of the static group membership feature that was implemented through several commits across multiple releases. Since then, we’re better prepared to skip releases if necessary.
Having an internal fork of Kafka where patches can be applied is tremendously helpful to our team because it does not prevent us from applying specific fixes to our environment.
After choosing a candidate release, we first look at the list of all issues filed against that version. Specifically, we consider blockers and critical, major bugs, and decide whether any of them are showstoppers. If so, there are a few options:
For our upgrade, we considered these bugs, and after confirming that the important ones for us were already fixed, we cherry-picked several fixes to create the upgrade candidate. Note that 2.3.2 was not available at the time.
Since we operate a mission-critical environment, it is prudent not to simply trust that a Kafka release would work well just because it is certified by the Kafka community. Once we select a candidate version, we try to create pipelines identical to our production pipelines. For us, this is often very easy because we can enable double publishing on the Singer side and the S3 Transporter for the test cluster. Once this is set up, we compare the important metrics (throughput, CPU usage, as well as data integrity validations) to certify our internal release. We’ve created an internal suite of validation checklists that can be repeatedly executed each time an upgrade is being evaluated.
We initially planned to upgrade to 2.4.1, but during our testing of this version, we noticed that the number of fetch requests unexpectedly spiked (almost doubled) after the upgrade, which caused the CPU usage of brokers to go up. The issue turned out to be a side effect of KIP-392 implementation and unnecessary watermark propagation in the default leader selector case. The fix for this issue was included in the 2.6.0 release. At Pinterest’s scale, this CPU spike was unacceptable due to its impact on resource usage (cost) and client SLAs; therefore, we had to change course and go with 2.3.1.
When upgrading clusters that are on a very old log message format version, we verify that none of the clients are on a client library version that would break once the log message format version is upgraded. If there are old clients like this, we make sure that those clients upgrade their Kafka client library version to a compatible version before upgrading the log message format version of the cluster. If, for some reason, certain clients cannot be upgraded, we do a log message format version override for the topics they work with so that the format version of those topics is not upgraded when the default format version is bumped. Also, note that for special cases/exceptions, we were able to isolate topics using brokersets that allowed us to provision the capacity needed to support message conversions as needed.
We went through a comprehensive exercise to understand how the log message format version upgrade would impact different client library versions of Java (native), C++ (librdkafka), Python (kafka-python and confluent-kafka-python), etc. This helped us understand whether any existing client would break and also whether any would lead to message format conversion when the default format version on the cluster is upgraded.
The following results show compatibility and also potential message conversion after upgrading log message format versions for Kafka version 2.0.0. Each row represents the results for a specific client library version, and each column represents the results for a specific log message format version.
The following is a key to the table below:
We are actively working on improving Logging Platform services at Pinterest and providing a better experience to our customers. The main focus areas at the moment include:
There are several other reasons why we started implementing a generic pub/sub client that abstracts the backend pub/sub system for consumers and addresses the gaps mentioned above. We believe this provides a win-win situation for the Logging Platform and its customers, because the Logging Platform will have more control and visibility into how clients interact with Kafka clusters. Furthermore, customers will not have to worry about issues and incidents when the pub/sub client library is at fault. It also enables us to provide other pub/sub system offerings in our environment when they are a better fit for specific use cases.
If you’d like to learn more, check out our Kafka Summit talk: Organic Growth and a Good Night’s Sleep: Effective Kafka Operations at Pinterest.
Our progress would not have been possible without tremendous contributions by Eric Lopez, Heng Zhang, Henry Cai, Jeff Xiang, and Ping-Min Lin.
Ambud Sharma is the tech lead and engineering manager for the Logging Platform team at Pinterest. Over the last two years, he has worked on architecting, stabilizing, and scaling the Logging Platform at Pinterest. Over the last five years, he has worked on building several petabyte-scale distributed systems at multiple Fortune 500 companies.