Category: DevOps

Migrating Elephants – How To Migrate Petabyte Scale Hadoop Clusters With Zero Downtime

 

Outbrain has been an early adopter of Hadoop and we, the team operating it, have acquired a lot of experience running it in production in terms of data ingestion, processing, monitoring, upgrading etc. This also means that we have a significant ecosystem around each cluster, with both open source and in-house systems.

 

A while back we decided to upgrade both the hardware and software versions of our Hadoop clusters.

“Why is that a big problem?” you might ask, so let me explain a bit about our current Hadoop architecture. We have two clusters of 300 machines in two different data centers, production and DR. Each cluster has a total dataset size of 1.5 PB with 5TB of compressed data loaded into it each day. There are ~10,000 job executions daily of about 1200 job definitions that were written by dozens of developers, data scientists and various other stakeholders within the company, spread across multiple teams around the globe. These jobs do everything from moving data into Hadoop (for ex. Sqoop or Mysql to Hive data loads), processing in Hadoop (for ex. running Hive, Scalding or Pig jobs), and pushing the results into external data stores (for ex. Vertica, Cassandra, Mysql etc.). An additional dimension of complexity originates from the dynamic nature of the system since developers, data scientists and researchers are pushing dozens of changes to how flows behave in production on a daily basis.

This system needed be migrated to run on new hardware, using new versions of multiple components of the Hadoop ecosystem, without impacting production processes and active users. A partial list of the components and technologies that are currently being used and should be taken into consideration is HDFS, Map-Reduce, Hive, Pig, Scalding and Sqoop. On top of that, of course, we have several more in-house services for data delivery, monitoring and retention that we have developed.

I’m sure you’ll agree that this is quite an elephant.

Storming Our Brains

We sat down with our users, and started thinking about a process to achieve this goal and quickly arrived at several guidelines that our selected process should abide by:

  1. Both Hadoop clusters (production and DR) should always be kept fully operational
  2. The migration process must be reversible
  3. Both value and risk should be incremental

After scratching our heads for quite a while, we came up with these options:

 

  1. In place: In place migration of the existing cluster to new version and then rolling the hardware upgrade by gradually pushing new machines into the cluster and removing the old machines. This is the simplest approach and you should probably have a very good reason to choose a different path if you can afford the risk. However since upgrading the system in place would expose clients to a huge change in an uncontrolled manner and is not by any means an easily reversible process we had to forego this option.
  2. Flipping the switch: The second option is to create a new cluster on new hardware, sync the required data, stop processing on the old cluster and move it to the new one. The problem here is that we still couldn’t manage the risk, because we would be stopping all processing and moving it to the new cluster. We wouldn’t know if the new cluster can handle the load or if each flow’s code is compatible with the new component’s version. As a matter of fact, there are a lot of unknowns that made it clear we had to split the problem into smaller pieces. The difficulty with splitting in this approach is that once you move a subset of the processing from the old cluster to the new, these results will no longer be accessible on the old cluster. This means that we would have had to migrate all dependencies of that initial subset. Since we have 1200 flow definitions with marvelous and beautiful interconnections between them, the task of splitting them would not have been practical and very quickly we found that we would have to migrate all flows together.
  3. Side by side execution: The 3rd option is to start processing on the new cluster without stopping the old cluster. This is a sort of an active-active approach, because both Hadoop clusters, new and old, will contain the processing results. This would allow us to migrate parts of the workload without risking interfering with any working pipeline in the old cluster. Sounds good, right?


  

First Steps

To better understand the chosen solution let’s take a look at our current architecture:

We have a framework that allows applications to push raw event data into multiple Hadoop clusters. For the sake of simplicity the diagram describes only one cluster. 

Once the data reaches Hadoop, processing begins to take place using a framework for orchestrating data flows we’ve developed in house that we like to call the Workflow Engine.

Each Workflow Engine belongs to a different business group. That Workflow Engine is responsible for triggering and orchestrating the execution of all flows developed and owned by that group. Each job execution can trigger more jobs on its current Workflow Engine or trigger jobs in other business groups’ Workflow Engines. We use this partitioning mainly for management and scale reasons but during the planning of the migration it provided us with a natural way to partition the workload, since there are very few dependencies between groups vs within each group.

 

Now that you have a better understanding of the existing layout you can see that the first step is to install a new Hadoop cluster with all required components of its ecosystem and begin pushing data into it.

To achieve this, we configured our dynamic data delivery pipeline system to send all events to the new cluster as well as the old, so now we have a new cluster with a fully operational data delivery pipeline:

 

Side by Side

Let’s think a bit about what options we had for running a side by side processing architecture.

We could use the same set of Workflow Engines to execute their jobs on both clusters, active and new. While this method would have the upside of saving machines and lower operational costs it would potentially double the load on each machine since jobs are assigned to machines in a static manner. This is due to the fact that each Workflow Engine is assigned a business group and all jobs that belong to this group are executed from it. To isolate the current production jobs execution from the ones for the new cluster we decided to allocate independent machines for the new cluster.

 

Let the Processing Commence!

Now that we have a fully operational Hadoop cluster running alongside our production cluster, and we now have raw data delivered into it, you might be tempted to say: “Great! Bring up a set of Workflow Engines and let’s start side by side processing!”.

 

Well… not really.

 

Since there are so many jobs and they doing varied types of operations we can’t really assume that letting them run side by side is a good idea. For instance, if a job calculates some results and then pushes them to MySql, these results will be pushed twice. Aside from doubling the load on the data bases for no good reason it may cause in some cases corruption or inconsistencies of the data due to race conditions. In essence, every job that writes to an external datasource should be allowed to run only once.

 

So we’ve described two types of execution modes a WorkflowEngine can have:

Leader: Run all the jobs!

Secondary: Run all jobs except those that might have a side effect external to that Hadoop cluster (e.g. write to external database or trigger an applicative service). This will be done automatically by the framework thus preventing any effort from the development teams.

 

When a Workflow Engine is in secondary mode, jobs executed from it can read from any source, but write only to a specific Hadoop cluster. That way they are essentially filling it up  and syncing (to a degree) with the other cluster.

 

Lets Do This…

Phase 1 of the migration should look something like this:

 

Notice that I’ve only included a Workflow Engine for one group in the diagram for simplicity but it will look similar for all other groups.

So the idea is to bring up a new Workflow Engine and give it the role of a migration secondary. This way it will run all jobs except for those writing to external data stores, thus eliminating all side effects external to the new Hadoop cluster.

By doing so, we were able to achieve multiple goals:

  1. Test basic software integration with the new Hadoop cluster version and all services of the ecosystem (hive, pig, scalding, etc.)
  2. Test new cluster’s hardware and performance compared to the currently active cluster
  3. Safely upgrade each business group’s Workflow Engine separately without impacting other groups.

 

Since the new cluster is running on new hardware and with a new version of Hadoop ecosystem, this is a huge milestone towards validating our new architecture. The fact the we managed to do so without risking any downtime that could have resulted from failing processing flows, wrong cluster configurations or any other potential issue was key in achieving our migration goals. 

 

Once we were confident that all phase 1 jobs were operating properly on the new cluster we could continue to phase 2 in which a migration leader becomes secondary and the secondary becomes a leader. Like this:

 

In this phase all jobs will begin running from the new Workflow Engine impacting all production systems, while the old Workflow Engine will only run jobs that create data to the old cluster. This method actually offers a fairly easy way to rollback to the old cluster in case of any serious failure (even after a few days or weeks) since all intermediate data will continue to be available on the old cluster.

 

The Overall Plan

The overall process is to push all Workflow Engines to phase 1 and then test and stabilize the system. We were able to run 70% (!) of our jobs in this phase. That’s 70% of our code, 70% of our integrations and APIs and at least 70% of the problems you would experience in a real live move. We were able to fix issues, analyze system performance and validate results. Only once everything seems to be working properly we can start pushing the groups to phase 2 one by one into a tested, stable new cluster.

Once again we benefit from the incremental nature of the process. Each business group can be pushed into phase 2 independently of other groups thus reducing risk and increasing our ability to debug and analyze issues. Additionally, each business group can start leveraging the new cluster’s capabilities (e.g. features from newer version, or improved performance) immediately after they have moved to phase 2 and not after we have migrated every one of the ~1200 jobs to run on the new cluster. One pain point that can’t be ignored is that inter-group dependencies can make this a significantly more complicated feat as you need to bring into consideration the state of multiple groups when migrating.

 

What Did We Achieve?

  1. Incremental Migration – Due to the fact that we had an active – active migration that we could apply on each business group, we benefited in terms of mitigating risk and gaining value from the new system gradually.
  2. Reversible process- since we kept all old workflowEngines (that executed their jobs on the old Hadoop cluster) in a state of secondary execution mode, all intermediate data was still being processed and was available in case we needed to revert groups independently from each other.
  3. Minimal impact on users – Since we defined an automated transition of jobs between secondary and leader modes users, didn’t need to duplicate any of their jobs.

 

What Now?

We have completed the upgrade and migration of our main cluster and have already started the migration of our DR cluster.

There are a lot more details and concerns to bring into account when migrating a production system at this scale. However, the basic abstractions we’ve introduced here, and the capabilities we’ve infused our systems with have equipped us with the tools to migrate elephants.

For more information about this project you can check out the video from Strata 2017 London where I discussed it in more detail.

Failure Testing for your private cloud – Introducing GomJabbar

Fork me on GitHub

TL;DR Chaos Drills can contribute a lot to your services resilience, and it’s actually quite a fun activity. We’ve built a tool called GomJabbar to help you run those drills.


Here at Outbrain we manage quite a large scale deployment of hundreds of services / modules, and thousands of hosts. We practice CI/CD, and implemented quite a sound infrastructure, which we believe is scalable, performant, and resilient. We do however experience many production issues on a daily basis, just like any other large scale organization. You simply can’t ensure a 100% fault free system. Servers will crash, run out of disk space, and lose connectivity to the network. Software will experience bugs, and erroneous conditions. Our job as software engineers is to anticipate these conditions, and design our code to handle them gracefully.

For quite a long time we were looking into ways of improving our resilience, and validate our assumptions, using a tool like Netflix’s Chaos Monkey. We also wanted to make sure our alerting system actually triggers when things go wrong. The main problem we were facing is that Chaos Monkey is a tool that was designed to work with cloud infrastructure, while we maintain our own private cloud.

The main motivation for developing such a tool, is that failures have the tendency of occurring when you’re least prepared, and in the least desirable time, e.g. Friday nights, when you’re out having a pint with your buddies. Now, to be honest with ourselves, when things fail during inconvenient times, we don’t always roll our sleeves and dive in to look for the root cause. Many times the incident will end after a service restart, and once the alerts clear we forget about it.

Wouldn’t it be great if we could have “chaos drills”, where we could practice handling failures, test and validate our assumptions, and learn how to improve our infrastructure?

Chaos Drills at Outbrain

We built GomJabbar exactly for the reasons specified above. Once a week, at a well known time, mid day, we randomly select a few targets where we trigger failures. At this point, the system should either auto-detect the failures, and auto-heal, or bypass them. In some cases alerts should be triggered to let teams know that a manual intervention is required.

After each chaos drill we conduct a quick take-in session for each of the triggered failures, and ask ourselves the following questions:

  1. Did the system handle the failure case correctly?
  2. Was our alerting strategy effective?
  3. Did the team have the knowledge to handle, and troubleshoot the failure?
  4. Was the issue investigated thoroughly?

These take-ins lead to super valuable inputs, which we probably wouldn’t collect any other way.

How did we kick this off?

Before we started running the chaos drills, there were a lot of concerns about the value of such drills, and the time it will require. Well, since eliminating our fear from production is one of the key goals of this activity, we had to take care of that first.

"I must not fear.
 Fear is the mind-killer.
 Fear is the little-death that brings total obliteration.
 I will face my fear.
 I will permit it to pass over me and through me.
 And when it has gone past I will turn the inner eye to see its path.
 Where the fear has gone there will be nothing. Only I will remain."

(Litany Against Fear - Frank Herbert - Dune)

So we started a series of chats with the teams, in order to understand what was bothering them, and found ways to mitigate it. So here goes:

  • There’s an obvious need to avoid unnecessary damage.
    • We’ve created filters to ensure only approved targets get to participate in the drills.
      This has a side effect of pre-marking areas in the code we need to take care of.
    • We currently schedule drills via statuspage.io, so teams know when to be ready, and if the time is inappropriate,
      we reschedule.
    • When we introduce a new kind of fault, we let everybody know, and explain what should they prepare for in advance.
    • We started out from minor faults like graceful shutdowns, continued to graceless shutdowns,
      and moved on to more interesting testing like faulty network emulation.
  • We’ve measured the time teams spent on these drills, and it turned out to be negligible.
    Most of the time was spent on preparations. For example ensuring we have proper alerting,
    and correct resilience features in the clients.
    This is actually something you need to do anyway. At the end of the day, we’ve heard no complaints about interruptions, nor time waste.
  • We’ve made sure teams, and engineers on call were not left on their own. We wanted everybody to learn
    from this drill, and when they were’nt sure how to proceed, we jumped in to help. It’s important
    to make everyone feel safe about this drill, and remind everybody that we only want to learn and improve.

All that said, it’s important to remember that we basically simulate failures that occur on a daily basis. It’s only that when we do that in a controlled manner, it’s easier to observe where are our blind spots, what knowledge are we lacking, and what we need to improve.

Our roadmap – What next?

  • Up until now, this drill was executed in a semi-automatic procedure. The next level is to let the teams run this drill on a fixed interval, at a well known time.
  • Add new kinds of failures, like disk space issues, power failures, etc.
  • So far, we were only brave enough to run this on applicative nodes, and there’s no reason to stop there. Data-stores, load-balancers, network switches, and the like are also on our radar in the near future.
  • Multi-target failure injection. For example, inject a failure to a percentage of the instances of some module in a random cluster. Yes, even a full cluster outage should be tested at some point, in case you were asking yourself.

The GomJabbar Internals

GomJabbar is basically an integration between a discovery system, a (fault) command execution scheduler, and your desired configuration. The configuration contains mostly the target filtering rules, and fault commands.

The fault commands are completely up to you. Out of the box we provide the following example commands, (but you can really write your own script to do what suits your platform, needs, and architecture):

  • Graceful shutdowns of service instances.
  • Graceless shutdowns of service instances.
  • Faulty Network Emulation (high latency, and packet-loss).

Upon startup, GomJabbar drills down via the discovery system, fetches the clusters, modules, and their instances, and passes each via the filters provided in the configuration files. This process is also performed periodically. We currently support discovery via consul, but adding other methods of discovery is quite trivial.

When a users wishes to trigger faults, GomJabbar selects a random target, and returns it to the user, along with a token that identifies this target. The user can then trigger one of the configured fault commands, or scripts, on the random target. At this point GomJabbar uses the configured CommandExecutor in order to execute the remote commands on the target hosts.

GomJabbar also maintains a audit log of all executions, which allows you to revert quickly in the face of a real production issue, or an unexpected catastrophe cause by this tool.

What have we learned so far?

If you’ve read so far, you may be asking yourself what’s in it for me? What kind of lessons can I learn from these drills?

We’ve actually found and fixed many issues by running these drills, and here’s what we can share:

  1. We had broken monitoring and alerting around the detection of the integrity of our production environment. We wanted to make sure that everything that runs in our data-centers is managed, and at a well known (version, health, etc). We’ve found that we didn’t compute the difference between the desired state, and the actual state properly, due to reliance on bogus data-sources. This sort of bug attacked us from two sides: once when we triggered graceful shutdowns, and once for graceless shutdowns.
  2. We’ve found services that had no owner, became obsolete, and were basically running unattended in production. The horror.
  3. During the faulty network emulations, we’ve found that we had clients that didn’t implement proper resilience features, and caused cascading failures in the consumers several layers up our service stack. We’ve also noticed that in some cases, the high latency also cascaded. This was fixed by adding proper timeouts, double-dispatch, and circuit-breakers.
  4. We’ve also found that these drills motivated developers to improve their knowledge about the metrics we expose, logs, and the troubleshooting tools we provide.

Conclusion

We’ve found the chaos drills to be an incredibly useful technique, which helps us improve our resilience and integrity, while helping everybody learn about how things work. We’re by no means anywhere near perfection. We’re actually pretty sure we’ll find many many more issues we need to take care of. We’re hoping this exciting new tool will help us move to the next level, and we hope you find it useful too 😉

I WANT IT ALL – Go Hybrid

When I was a kid, my parents used to tell me that I can’t have my cake and eat too.  Actually that’s a lie, they never said that. Still it is something I hear parents say quite often. And not just parents. I meet the same phrase everywhere I go. People constantly taking a firm, almost religious stance about choosing one thing over another: Mac vs PC, Android vs iOS, Chocolate vs Vanilla (obviously Chocolate!).

So I’d like to take a moment to take a different, more inclusive approach.

Forget Mac vs PC. Forget Chocolate vs Vanilla.

I don’t want to choose. I Want it all!

At Outbrain, the core of our compute infrastructure is based on bare metal servers. With a fleet of over 6000 physical nodes, spread across 3 datacenters, we’ve learned over the years how to manage an efficient, tailored environment that caters to our unique needs. One of which being the processing and serving of over 250 Billion personalized recommendations a month, to over 550 Million unique users.

Still, we cannot deny that the Cloud brings forth advantages that are hard to achieve in bare metal environments. And in the spirit of inclusiveness (and maximising value), we want to leverage these advantages to complement and extend what we’ve already built. Whether focusing on workloads that require a high level of elasticity, such as ad-hoc research projects involving large amount of data, or simply external services that can increase our productivity. We’ve come to view Cloud Solutions as supplemental to our tailored infrastructure rather than a replacement.

 

Over recent months, we’ve been experimenting with 3 different vectors involving the Cloud:

 

Elasticity

Our world revolves around publications, especially news. As such, whenever a major news event occurs, we feel immediate, potentially high impact. Users rush to publisher sites, where we are installed. They want their news, they want their recommendations, and they want them all now.

For example, when Carrie Fisher, AKA Princess Leia, passed away last December, we saw a 30% traffic increase on top of our usual peak traffic. That’s quite a spike.

Since usually we do not know when the breaking news event will be, it means that we are required to keep enough extra capacity to support such surges.

By leveraging the cloud, we can keep that additional extra capacity to bare minimum, relying instead on the inherent elasticity of the cloud, provisioning only what we need when we need it.

Doing this can improve the efficiency of our environment and cost model.

Ad-hoc Projects

A couple of months back one of researchers came up with an interesting behavioral hypothesis. For the discussion at hand, lets say that it was “people who like chocolate are more likely to raise pet gerbils.” (drop a comment with the word “gerbils” to let me know that you’ve read thus far). That sounded interesting, but raised a challenge. To validate or disprove this, we needed to analyze over 600 Terabytes of data.

We could have run it on our internal Hadoop environment, but that came with a not-so-trivial price tag. Not only did we have to provision additional capacity in our Hadoop cluster to support the workload, we anticipated the analysis to also carry impact on existing workloads running in the cluster. And all this before getting into operational aspects such as labor and lead time.

Instead, we chose to upload the data into Google’s BigQuery. This gave us both shorter lead times for the setup and very nice performance. In addition, 3 months into the project, when the analysis was completed, we simply shut down the environment and were done with it. As simple as that!

Productivity

We use Fastly for dynamic content acceleration. Given the scale we mentioned, this has the side-effect of generating about 15 Terabytes of Fastly access logs each month. For us, there’s a lot of interesting information in those logs. And so, we had 3 alternatives when deciding how to analyse them:

  •      SaaS based log analysis vendors
  •      An internal solution, based on the ELK stack
  •      A cloud based solution, based on BigQuery and DataStudio

After performing a PoC and running the numbers, we found that the BigQuery option – if done right – was the most effective for us. Both in terms of cost, and amount of required effort.

There are challenges when designing and running a hybrid environment. For example, you have to make sure you have consolidated tools to manage both on-prem and Cloud resources. The predictability of your monthly cost isn’t as trivial as before (no one likes surprises there!), controls around data can demand substantial investments… but that doesn’t make the fallback to “all Vanilla” or “all Chocolate” a good one. It just means that you need to be mindful and prepared to invest in tooling, education and processes.

 

In summary, I’d like to revisit my parents advice, and try to improve on it a bit (which I’m sure they won’t mind!):

Be curious. Check out what is out there. If you like what you see – try it out. At worst, you’ll learn something new. At best, you’ll have your cake… and eat it too.