Learn how Wallaroo helped Pubmatic with infrastructure costs and time-to-market.

Learn more

The Treacherous Tangle of Redundant Data: Resilience for Wallaroo

Introduction: we need data redundancy, but how, exactly?

You now have your distributed system in production, congratulations! Your cluster is starting at six machines, but it is expected to grow quickly as it is assigned more work. The cluster’s main application is stateful, and that’s a problem. What if you lose a local disk drive? Or a sysadmin runs rm -rf on the wrong directory? Or else the entire machine cannot reboot, due to a power failure or administrator error that destroys an entire virtual machine?

Today, your application is writing all of its critical data to the files in the local file system of each machine. Tomorrow, that data needs to be redundant. Sooner or later, your hardware will fail or an administrator will do something wrong.

How should you add the data redundancy that we know that you need? You ask for advice, and the answers you get are scattered all over the map of tech.

  • Store the files with an NFS NAS or perhaps SAN or iSCSI service or an AWS Elastic Block Store (EBS) volume

  • Store the files in non-traditional shared file service like Hadoop or Ceph

  • Switch from files to a relational database like replicated MySQL or PostgreSQL or Azure’s CosmosDB

  • Switch from files to a non-relational database like Cassandra or Amazon’s DynamoDB

  • None of the above

Each of those answers could make sense, depending on the application and your company and the company’s broader goals. This article explores how Wallaroo Labs chose a data redundancy technique that fits the needs of our product, Wallaroo.

What is Wallaroo?

Wallaroo is a framework designed to make it easier for developers to build and operate high-performance applications. It handles the complexity of building distributed data processing applications so all that you need to worry about is the domain logic.

Wallaroo is a streaming data processing framework. One property of “streaming data” is that the total size of the input is unknown. The last item to process may never arrive. Streaming platforms also don’t know how much data will arrive in any given timeframe. How many events will arrive in the next minute? Zero? One thousand? Ninety million? Every minute, you aren’t certain when or how much data will arrive for processing.

Options for partial failure in distributed stream processing

When a machine in a distributed stream processor fails, there are two basic options for recovery:

  1. Start over. That is, process all stream data again, starting from the beginning.
  2. Restore system state from some snapshot or backup that was made in the recent past, then resume processing.

Option 1 is what streaming systems based on Apache Spark and Kafka do: they store all the input event data, so they can resend the events after a component has rebooted.

Option 2 is what Wallaroo does. My colleague, John Mumm, recently wrote a blog article about Wallaroo’s checkpointing algorithm. That article explains Wallaroo’s recovery process when a worker can be restarted. However, that blog post doesn’t cover a different scenario: what happens if a Wallaroo worker’s data is lost?

When data is lost forever, by example

Let’s look at a concrete example: a Wallaroo cluster with six worker processes, each on a different computer (or virtual machine or container).

Six worker Wallaroo cluster, state storage via local disk Figure 1: Six worker Wallaroo cluster, state storage via local disk

Figure 1 has no data redundancy in it. If any local disk fails, then recovery is impossible. If an entire machine or VM or container fails by itself or by administrator error, then recovery is impossible.

As an alternative, what if the same cluster were deployed in an Amazon Web Services (AWS) cluster using Elastic Block Store (EBS) volumes to create data redundancy? That’s what Figure 2 depicts. The same general principle applies to most file- block-device-centric schemes mentioned in the introduction, including NAS, SAN, iSCSI, and so on.

Six worker Wallaroo cluster, state storage via EBS Figure 2: Six worker Wallaroo cluster, state storage via EBS

Let’s assume that worker5’s machine caught fire and its data is lost; see Figure 3. Also, let’s make the diagram a bit more accurate than Figure 2’s placement of the EBS service relative to the rest of the cluster.

Wallaroo worker5 process restarts on a replacement machine/VM/container, accessing an EBS volume Figure 3: Wallaroo worker5 process restarts on a replacement machine/VM/container, accessing an EBS volume

When the original machine/VM fails, the EBS volume is disconnected from the dead VM and attached to a replacement machine/VM. Then worker5 is restarted on that machine, and worker5 participates in Wallaroo’s internal snapshot recovery protocol to finish the recovery process.

Your Boss Says, “Sorry, What About …?”

At some point, your boss or perhaps you, will start adding some annoying restrictions to your app.

  • We can’t use Amazon AWS exclusively for this app.
  • We can’t use our on-premise data center for this app.
  • The EMC SAN is being phased out soon, so this app cannot use it.

This is just a small sample of the deployment restrictions you might have to work with. Let’s look at the properties, restrictions, and priorities that Wallaroo Labs considered when creating a data redundancy strategy for Wallaroo.

Recovery data is infinitely sized and arrives at unpredictable times

Wallaroo is a streaming data system. Two fundamental properties of a streaming data system are the following:

  1. The input data is infinite. Truly infinite data is impossible, but the input data may be very, very large. Many network data protocols assume that the amount of data to transfer is known in advance. An ideal data replication system would not require Wallaroo to specify exact data sizes in advance, for example, as Amazon’s S3 object store requires. (The multi-part upload API is more complicated and moves the problem to a different place.)

  2. The input data’s arrival time is unknown. For every data streaming system that receives at least one byte per second, there is a system that may receive only one byte every day or every week. With large gaps between data input events, Wallaroo could run into data replication problems related to timeouts. When writing data to an FTP server or Amazon’s S3 service or to an SQL server, when does the server disconnect due to an idle/timeout event?

Append-only files by single producer, single consumer

Wallaroo has some properties that simplify choosing a redundant data storage scheme: a single producer and a single consumer.

General purpose databases and file systems usually support all kinds of shared read and write access to their data. If 700 clients want to read and write to the same database row or key or file, they will support it.

In contrast, every Wallaroo worker’s recovery data is not shared with other workers: it is private state, for use by a single worker only. A single worker process only runs on one machine at any given time. The worker’s recovery data is accessed in a single producer and single consumer manner. Industry and open source communities have many solutions for this kind of problem, including these types:

  • Storage Area Networks (SANs) such as Fibre Channel shared disks, FCoE (Fibre Channel over Ethernet), and AoE (ATA over Ethernet)

  • Exclusive use shared block devices such as AWS’s EBS (Elastic Block Store) and Linux’s NBD (Network Block Device).

However, most of these techniques make assumptions about hardware and/or network services that are not available in all public cloud computing service providers or in private data centers. We do not want Wallaroo’s data redundancy management to steer us into a vendor lock-in problem.

Furthermore, most of Wallaroo worker’s data files are append-only. We very intentionally made this choice to make file replication easier. An append-only file’s data, once it’s written, is not modified by later writes. Once written, an append-only file’s contents are immutable. This immutability makes the task of replicating the file much, much easier. (For further reading, see Pat Helland’s excellent article with the punny name, Immutability Changes Everything.)

Reduce complexity as much as feasible

There are many open source systems today that can support Wallaroo’s recovery data storage requirements. Apache Kafka and Bookkeeper are two of the best-known ones. Wallaroo’s use case fits Kafka’s and Bookkeeper’s feature sets almost exactly. Why not use Kafka? Or Bookkeeper? Hadoop is another excellent fit for Wallaroo’s needs. Why not use Hadoop?

Alas, Kafka, Bookkeeper, and Hadoop are complex pieces of software in their own right. Wallaroo does not need most of their features. Let’s try to avoid requiring that Wallaroo administrators be familiar with details of running those services in development, testing, and production environments.

Kafka, Bookkeeper, and Hadoop have another complex dependency: the ZooKeeper consensus system. Nobody at Wallaroo Labs wishes to make ZooKeeper mandatory for development work or for production systems. Let’s try to avoid requiring that our customers use and maintain ZooKeeper.

Avoid vendor lock-in

Wallaroo Labs does not want to restrict Wallaroo deployments to a single cloud provider like Azure, Google Compute Engine, or Amazon Web Services. We could avoid lock-in by supporting many or all such cloud services. However, a company of our size doesn’t have the resources to quickly develop the code needed for supporting all major players in the cloud storage market.

Similarly, for traditional or on-premise data centers, we didn’t want to require the use of a NAS, SAN, or similar storage device.

Almost good enough: FTP, the File Transfer Protocol from the 1980s

RFC 959 defines the File Transfer Protocol (FTP). FTP services are old, well-tested, and are widely available. FTP, however, is not a perfect fit for Wallaroo’s use case because of the following conditions:

  1. Wallaroo’s data arrives at unpredictable times. We could tweak an FTP server to use infinite timeouts, if we wished.

  2. A single writer restriction is not available. Again, we could tweak an FTP server to enforce this restriction, if we wished.

  3. Periodic feedback about recevied data is not available. Adding feedback for received and fsync’ed data (e.g., once per second) would be a significant protocol change, not a simple tweak. FTP provides no status acknowledgment until the end of file is received. Also, RFC 959 does not explicitly require durable or stable storage of a file’s contents.

Wallaroo’s choice: DOS, the Dumb Object Service, that is almost FTP

The last section listed three problems with the FTP protocol as defined by RFC 959. For each problem, a possible “tweak” was mentioned. That was not an accident. If the FTP protocol were tweaked in those three ways, then the result would fit Wallaroo’s needs exactly.

Rather than adapt the FTP protocol, we chose to create and implement a new file transfer protocol that is very similar to FTP but not quite FTP. Its name is DOS, the Dumb Object Service.

The DOS protocol is inspired by a small subset of FTP’s commands but without the ASCII-only nature of the FTP commands. In a very small binary protocol, we specify these FTP-like commands:

  • CWD: change the working directory. This feature is used to allow a single DOS server to store data for multiple DOS clients: each DOS client writes its data files into a separate directory.

  • APPEND: append data to a file. It is more similar to HTTP’s PUT command than to the FTP command combination of PASV and STOR.

The APPEND command includes an argument for the DOS server’s file size (or size 0 if the file does not yet exist). The DOS server will allow the APPEND operation to proceed only if the server’s file size is exactly equal to the size argument provided by the DOS client.

Also, the server prohibits multiple clients writing to the same file simultaneously.

  • GET: retrieve a file’s contents. It is more similar to HTTP’s GET command than to the FTP command combination of PASV and RETR. The DOS server’s GET command permits partial reads of the file by specifying a starting offset and maximum number of bytes to send to the client.

  • LIST: list files in the current directory. This feature isn’t mentioned in the requirements list above, but it is useful for a Wallaroo worker client to discover the names and sizes of files stored in its own private directory.

  • QUIT: terminates a DOS client session.

Examples of Wallaroo clusters with DOS servers providing data redundancy

Let’s see what happens when we add two DOS servers into our six-worker Wallaroo cluster. Each worker is configured to replicate its I/O journal files to both DOS servers. Each DOS server uses its local file system and local disk(s) to store the I/O journals for all six Wallaroo workers.

Six worker Wallaroo cluster plus two DOS servers Figure 4: Six worker Wallaroo cluster plus two DOS servers

When a Wallaroo worker fails catastrophically, its local recovery files are lost. We must retrieve copies from a remote DOS server before we can restart the worker. The procedure for figuring out which DOS server is most “up to date” is described later in this article.

Figure 4 suggests that 8 different machines (or virtual machines or containers) are required: six for Wallaroo worker processes and two for the DOS servers. Figure 5 shows what happens when we collocate the DOS servers with the boxes that also run two of the Wallaroo workers.

Six worker Wallaroo cluster plus two DOS servers sharing hardware Figure 5: Six worker Wallaroo cluster plus two DOS servers sharing hardware

Our last example is another option: it uses EBS to provide redundant storage for a single DOS server instead of each of the six worker servers in Figure 2. The system in Figure 6 may be a more cost-effective option if “provisioned IOPS” costs for each EBS volume are added to the system. When the DOS server fails, the replacement procedure is similar to the failover scenario shown in Figure 3.

Six worker Wallaroo cluster plus one DOS server plus EBS Figure 6: Six worker Wallaroo cluster plus one DOS server plus EBS

Implementation overview: client side (inside Wallaroo)

Each Wallaroo worker creates a number of recovery data files inside the --resilience-dir directory specified on the command line. A file listing of that directory looks like this:

-rw-r--r-- 1 ubuntu ubuntu   48 Oct  3 19:34 my_app-worker3.checkpoint_ids
-rw-r--r-- 1 ubuntu ubuntu  803 Oct  3 19:34 my_app-worker3.connection-addresses
-rw-r--r-- 1 ubuntu ubuntu  447 Oct  3 19:34 my_app-worker3.evlog
-rw-r--r-- 1 ubuntu ubuntu    0 Oct  3 19:34 my_app-worker3.local-keys
-rw-r--r-- 1 ubuntu ubuntu 4144 Oct  3 19:34 my_app-worker3.local-topology
-rw-r--r-- 1 ubuntu ubuntu   18 Oct  3 19:34 my_app-worker3.tcp-control
-rw-r--r-- 1 ubuntu ubuntu   18 Oct  3 19:34 my_app-worker3.tcp-data
-rw-r--r-- 1 ubuntu ubuntu   20 Oct  3 19:34 my_app-worker3.workers

The Wallaroo application has been intentionally restricted to only three file-mutating I/O operations: append file data, file truncate, and file delete. For each file I/O operation, the I/O call and its arguments are written to a journal file. All I/O to a journal file is strictly append-only. All journal file data is written via the DOS protocol to one or more DOS servers. I describe the serialization of these I/O operations in the next section.

Serialized I/O operations in the my_app.journal file

In the my_app-worker3.journal file, each file write/append event to a non-journal file (e.g., the my_app-worker3.checkpoint_ids file) is serialized as if a single generic function had the Pony type:

encode_request(optag: USize, op: U8,
    ints: Array[USize], bss: Array[ByteSeq]):
    (Array[ByteSeq] iso^, USize)

A request has an operation tag number, an operation specifying byte, and two arrays of arguments: one of the Pony language’s type USize for integers and one of Pony’s type ByteSeq for strings/byte arrays. Depending on the operation byte op, the caller includes the appropriate number of elements in each respective array.

A view of serialized bytes returned by encode_request is:

| Size | Type           | Description                     |
|    1 | U8             | Protocol version = 0            |
|    1 | U8             | Op type: append, delete, ...    |
|    1 | U8             | Number of int args              |
|    1 | U8             | Number of string/byte args      |
|    8 | USize          | Op tag                          |
|    8 | USize          | 1st int arg                     |
|    8 | USize          | nth int arg                     |
|    8 | USize          | Size of 1st string/byte arg     |
|    8 | USize          | Size of nth string/byte arg     |
|    X | String/ByteSeq | Contents of 1st string/byte arg |
|    X | String/ByteSeq | Contents of nth string/byte arg |

If the binary contents of the my_app.journal file were translated into a human-readable form, they look something like this:

version=0 op=1 tag=1 ints=[0] strings=['/w/my_app-worker3.tcp-control', 'localhost\n']
version=0 op=1 tag=2 ints=[10] strings=['/w/my_app-worker3.tcp-control', '0\n']
version=0 op=0 tag=3 ints=[0] strings=['/w/my_app-worker3.connection-addresses']
  • Where op=1 is a file write operation:
    • The first integer in the ints list is the file offset
    • The first string in the strings list is the path of the file to write to.
    • The second string in the strings list is the data written at this offset.
  • Where op=0 is a file truncate operation:
    • The first integer in the ints list is the file size to truncate to.
    • The first string in the strings list is the path of the file to truncate.

I/O Journal + Append-Only Journal I/O -> Journal File Size as Logical Time

All file-mutating I/O operations to --resilience-dir files are serialized and appended to a local I/O journal file. Asynchronously, the local I/O journal file data is written via the DOS protocol to one or more remote DOS servers for redundancy.

We can consider the file size of the I/O journal files as “logical time”. Given the set F of all files whose I/O ops are written in an I/O journal, then any single file f in the set F can be reconstructed at some logical time t by extracting all of f’s I/O events from the journal where the file offset of the I/O event is less than t.

For example, if we want to reconstruct (or “roll back”) the contents of the /w/my_app-worker3.connection-addresses file as it existed at the wall clock time of midnight on New Year’s Day 2018, the procedure is:

  1. Determine the size of the I/O journal file at midnight on New Year’s Day 2018. Let’s call this offset max_offset.
  2. Open the I/O journal file for reading and seek to the beginning of the journal.
  3. Read each serialized I/O operation in the file until we reach the offset max_offset.
  4. If a serialized operation involves the file path /w/my_app-worker3.connection-addresses, then apply the op (i.e., write/file truncate/file delete) to the file path.

Chandy-Lamport snapshot consistency across Wallaroo workers

Any logical time construct mentioned in this article is applicable only to a single Wallaroo worker and is not valid for comparison with other Wallaroo workers.

Wallaroo release 0.5.3 introduced a snapshot algorithm responsible for managing logical time consistency across all Wallaroo workers. That algorithm, a variation of the Chandy-Lamport snapshot algorithm, uses a separate notion of logical time than the logical time discussed here. Please see the article Checkpointing and Consistent Recovery Lines: How We Handle Failure in Wallaroo for an overview of how the Wallaroo’s adaptation of Chandy-Lamport works.

Assume no Byzantine failures or data corruption inside I/O journal files

The DOS protocol and larger Wallaroo system cannot operate correctly with Byzantine failures. We assume that there are no Byzantine failures, including software bugs that would do Bad Things™ such as reorder records within a journal. And we assume that there’s no data corruption (for any reason) inside of a journal file.

In the case of data corruption inside of the log, we do have remedies available. The best remedy is to add robust checksums to all data written to a journal. Today, Wallaroo does not yet use checksums.

What’s next for data redundancy and Wallaroo?

Wallaroo version 0.5.4 (released on October 31, 2018) now has all of the raw components for data redundancy. To date, those components are only lightly tested for performance. The DOS client, written purely in Pony inside of Wallaroo, has not had any performance analysis performed yet.

The DOS server is written in Python. That seems an odd choice for a high-performance server, but early tests have shown that that the DOS server can write over 100 Mbytes/second of recovery data without significant latency. A later Wallaroo release may reimplement the DOS server in Pony for higher efficiency.

The long-term goal is to have a hands-free, turnkey system that will react to the failure of any Wallaroo component and restart it gracefully. We need to choose the environment for such a system: Kubernetes or Mesos? GCE or Azure or AWS? “On-prem” data centers? All of the above? While we wrestle with these choices, we’re confident that we are building on a solid foundation of data redundancy components.

Give it a try!

Hopefully, I’ve gotten you excited about Wallaroo and our data redundancy story. We’d love for you to give it a try and give us your feedback. Your feedback helps us drive the product forward. Thank you to everyone who has contributed feedback so far, and thank you to everyone who does after reading this blog post.

Get started with Wallaroo now!

Did you enjoy this post?

Subscribe to our blog and be updated each time we add a new post.

Other posts you might also like: