Developing Storage Solutions Before the Rest with AB Periasamay

Episode Summary

Conversations about what the cloud is might be an infinitely convoluted one, but some are taking the conversation down paths less traveled. That is certainly the case for AB Periasamy, CEO and Co-Founder of MinIO, an open source provider of high performance, object storage software. AB and Corey talk about the shift from legacy to the cloud and the advent of S3, and how in that transition AB saw an opportunity. That was when he started to build a brand around the new demands for storage. AB saw that building a software storage company early on that was compatible with S3 would “sweep the market” and now they’re sitting on a billion dollar company. Corey and AB take a deeply technical dive, check it out!

Episode Show Notes & Transcript

About AB
AB Periasamy is the co-founder and CEO of MinIO, an open source provider of high performance, object storage software. In addition to this role, AB is an active investor and advisor to a wide range of technology companies, from H2O.ai and Manetu where he serves on the board to advisor or investor roles with Humio, Isovalent, Starburst, Yugabyte, Tetrate, Postman, Storj, Procurify, and Helpshift. Successful exits include Gitter.im (Gitlab), Treasure Data (ARM) and Fastor (SMART).

AB co-founded Gluster in 2005 to commoditize scalable storage systems. As CTO, he was the primary architect and strategist for the development of the Gluster file system, a pioneer in software defined storage. After the company was acquired by Red Hat in 2011, AB joined Red Hat’s Office of the CTO. Prior to Gluster, AB was CTO of California Digital Corporation, where his work led to scaling of the commodity cluster computing to supercomputing class performance. His work there resulted in the development of Lawrence Livermore Laboratory’s “Thunder” code, which, at the time was the second fastest in the world.  

AB holds a Computer Science Engineering degree from Annamalai University, Tamil Nadu, India.

AB is one of the leading proponents and thinkers on the subject of open source software - articulating the difference between the philosophy and business model. An active contributor to a number of open source projects, he is a board member of India's Free Software Foundation.


Links:

Transcript
Announcer: Hello, and welcome to Screaming in the Cloud with your host, Chief Cloud Economist at The Duckbill Group, Corey Quinn. This weekly show features conversations with people doing interesting work in the world of cloud, thoughtful commentary on the state of the technical world, and ridiculous titles for which Corey refuses to apologize. This is Screaming in the Cloud.


Corey: This episode is sponsored in part by our friends at Sysdig. Sysdig is the solution for securing DevOps. They have a blog post that went up recently about how an insecure AWS Lambda function could be used as a pivot point to get access into your environment. They’ve also gone deep in-depth with a bunch of other approaches to how DevOps and security are inextricably linked. To learn more, visit sysdig.com and tell them I sent you. That’s S-Y-S-D-I-G dot com. My thanks to them for their continued support of this ridiculous nonsense.


Corey: This episode is sponsored in part by our friends at Rising Cloud, which I hadn’t heard of before, but they’re doing something vaguely interesting here. They are using AI, which is usually where my eyes glaze over and I lose attention, but they’re using it to help developers be more efficient by reducing repetitive tasks. So, the idea being that you can run stateless things without having to worry about scaling, placement, et cetera, and the rest. They claim significant cost savings, and they’re able to wind up taking what you’re running as it is, in AWS, with no changes, and run it inside of their data centers that span multiple regions. I’m somewhat skeptical, but their customers seem to really like them, so that’s one of those areas where I really have a hard time being too snarky about it because when you solve a customer’s problem, and they get out there in public and say, “We’re solving a problem,” it’s very hard to snark about that. Multus Medical, Construx.ai, and Stax have seen significant results by using them, and it’s worth exploring. So, if you’re looking for a smarter, faster, cheaper alternative to EC2, Lambda, or batch, consider checking them out. Visit risingcloud.com/benefits. That’s risingcloud.com/benefits, and be sure to tell them that I said you because watching people wince when you mention my name is one of the guilty pleasures of listening to this podcast.in a silo


Corey: Welcome to Screaming in the Cloud. I’m Corey Quinn. I’m joined this week by someone who’s doing something a bit off the beaten path when we talk about cloud. I’ve often said that S3 is sort of a modern wonder of the world. It was the first AWS service brought into general availability. Today’s promoted guest is the co-founder and CEO of MinIO, Anand Babu Periasamy, or AB as he often goes, depending upon who’s talking to him. Thank you so much for taking the time to speak with me today.


AB: It’s wonderful to be here, Corey. Thank you for having me.


Corey: So, I want to start with the obvious thing, where you take a look at what is the cloud and you can talk about AWS’s ridiculous high-level managed services, like Amazon Chime. Great, we all see how that plays out. And those are the higher-level offerings, ideally aimed at problems customers have, but then they also have the baseline building blocks services, and it’s hard to think of a more baseline building block than an object store. That’s something every cloud provider has, regardless of how many scare quotes there are around the word cloud; everyone offers the object store. And your solution is to look at this and say, “Ah, that’s a market ripe for disruption. We’re going to build through an open-source community software that emulates an object store.” I would be sitting here, more or less poking fun at the idea except for the fact that you’re a billion-dollar company now.


AB: Yeah.


Corey: How did you get here?


AB: So, when we started, right, we did not actually think about cloud that way, right? “Cloud, it’s a hot trend, and let’s go disrupt is like that. It will lead to a lot of opportunity.” Certainly, it’s true, it lead to the M&S, right, but that’s not how we looked at it, right? It’s a bad idea to build startups for M&A.


When we looked at the problem, when we got back into this—my previous background, some may not know that it’s actually a distributed file system background in the open-source space.


Corey: Yeah, you were one of the co-founders of Gluster—


AB: Yeah.


Corey: —which I have only begrudgingly forgiven you. But please continue.


AB: [laugh]. And back then we got the idea right, but the timing was wrong. And I had—while the data was beginning to grow at a crazy rate, end of the day, GlusterFS has to still look like an FS, it has to look like a file system like NetApp or EMC, and it was hugely limiting what we can do with it. The biggest problem for me was legacy systems. I have to build a modern system that is compatible with a legacy architecture, you cannot innovate.


And that is where when Amazon introduced S3, back then, like, when S3 came, cloud was not big at all, right? When I look at it, the most important message of the cloud was Amazon basically threw everything that is legacy. It’s not [iSCSI 00:03:21] as a Service; it’s not even FTP as a Service, right? They came up with a simple, RESTful API to store your blobs, whether it’s JavaScript, Android, iOS, or [AAML 00:03:30] application, or even Snowflake-type application.


Corey: Oh, we spent ten years rewriting our apps to speak object store, and then they released EFS, which is NFS in the cloud. It’s—


AB: Yeah.


Corey: —I didn’t realize I could have just been stubborn and waited, and the whole problem would solve itself. But here we are. You’re quite right.


AB: Yeah. And even EFS and EBS are more for legacy stock can come in, buy some time, but that’s not how you should stay on AWS, right? When Amazon did that, for me, that was the opportunity. I saw that… while world is going to continue to produce lots and lots of data, if I built a brand around that, I’m not going to go wrong.


The problem is data at scale. And what do I do there? The opportunity I saw was, Amazon solved one of the largest problems for a long time. All the legacy systems, legacy protocols, they convinced the industry, throw them away and then start all over from scratch with the new API. While it’s not compatible, it’s not standard, it is ridiculously simple compared to anything else.


No fstabs, no [unintelligible 00:04:27], no [root 00:04:28], nothing, right? From any application anywhere you can access was a big deal. When I saw that, I was like, “Thank you Amazon.” And I also knew Amazon would convince the industry that rewriting their application is going to be better and faster and cheaper than retrofitting legacy applications.


Corey: I wonder how much that’s retconned because talking to some of the people involved in the early days, they were not at all convinced they [laugh] would be able to convince the industry to do this.


AB: Actually, if you talk to the analyst reporters, the IDC’s, Gartner’s of the world to the enterprise IT, the VMware community, they would say, “Hell no.” But if you talk to the actual application developers, data infrastructure, data architects, the actual consumers of data, for them, it was so obvious. They actually did not know how to write an fstab. The iSCSI and NFS, you can’t even access across the internet, and the modern applications, they ran across the globe, in JavaScript, and all kinds of apps on the device. From [Snap 00:05:21] to Snowflake, today is built on object store. It was more natural for the applications team, but not from the infrastructure team. So, who you asked that mattered.


But nevertheless, Amazon convinced the rest of the world, and our bet was that if this is going to be the future, then this is also our opportunity. S3 is going to be limited because it only runs inside AWS. Bulk of the world’s data is produced everywhere and only a tiny fraction will go to AWS. And where will the rest of the data go? Not SAN, NAS, HDFS, or other blob store, Azure Blob, or GCS; it’s not going to be fragmented. And if we built a better object store, lightweight, faster, simpler, but fully compatible with S3 API, we can sweep and consolidate the market. And that’s what happened.


Corey: And there is a lot of validity to that. We take a look across the industry, when we look at various standards—I mean, one of the big problems with multi-cloud in many respects is the APIs are not quite similar enough. And worse, the failure patterns are very different, of I don’t just need to know how the load balancer works, I need to know how it breaks so I can detect and plan for that. And then you’ve got the whole identity problem as well, where you’re trying to manage across different frames of reference as you go between providers, and leads to a bit of a mess. What is it that makes MinIO something that has been not just something that has endured since it was created, but clearly been thriving?


AB: The real reason, actually is not the multi-cloud compatibility, all that, right? Like, while today, it is a big deal for the users because the deployments have grown into 10-plus petabytes, and now the infrastructure team is taking it over and consolidating across the enterprise, so now they are talking about which key management server for storing the encrypted keys, which key management server should I talk to? Look at AWS, Google, or Azure, everyone has their own proprietary API. Outside they, have [YAML2 00:07:18], HashiCorp Vault, and, like, there is no standard here. It is supposed to be a [KMIP 00:07:23] standard, but in reality, it is not. Even different versions of Vault, there are incompatibilities for us.


That is where—like from Key Management Server, Identity Management Server, right, like, everything that you speak around, how do you talk to different ecosystem? That, actually, MinIO provides connectors; having the large ecosystem support and large community, we are able to address all that. Once you bring MinIO into your application stack like you would bring Elasticsearch or MongoDB or anything else as a container, your application stack is just a Kubernetes YAML file, and you roll it out on any cloud, it becomes easier for them, they’re able to go to any cloud they want. But the real reason why it succeeded was not that. They actually wrote their applications as containers on Minikube, then they will push it on a CI/CD environment.


They never wrote code on EC2 or ECS writing objects on S3, and they don’t like the idea of [past 00:08:15], where someone is telling you just—like you saw Google App Engine never took off, right? They liked the idea, here are my building blocks. And then I would stitch them together and build my application. We were part of their application development since early days, and when the application matured, it was hard to remove. It is very much like Microsoft Windows when it grew, even though the desktop was Microsoft Windows Server was NetWare, NetWare lost the game, right?


We got the ecosystem, and it was actually developer productivity, convenience, that really helped. The simplicity of MinIO, today, they are arguing that deploying MinIO inside AWS is easier through their YAML and containers than going to AWS Console and figuring out how to do it.


Corey: As you take a look at how customers are adopting this, it’s clear that there is some shift in this because I could see the story for something like MinIO making an awful lot of sense in a data center environment because otherwise, it’s, “Great. I need to make this app work with my SAN as well as an object store.” And that’s sort of a non-starter for obvious reasons. But now you’re available through cloud marketplaces directly.


AB: Yeah.


Corey: How are you seeing adoption patterns and interactions from customers changing as the industry continues to evolve?


AB: Yeah, actually, that is how my thinking was when I started. If you are inside AWS, I would myself tell them that why don’t use AWS S3? And it made a lot of sense if it’s on a colo or your own infrastructure, then there is an object store. It even made a lot of sense if you are deploying on Google Cloud, Azure, Alibaba Cloud, Oracle Cloud, it made a lot of sense because you wanted an S3 compatible object store. Inside AWS, why would you do it, if there is AWS S3?


Nowadays, I hear funny arguments, too. They like, “Oh, I didn’t know that I could use S3. Is S3 MinIO compatible?” Because they will be like, “It came along with the GitLab or GitHub Enterprise, a part of the application stack.” They didn’t even know that they could actually switch it over.


And otherwise, most of the time, they developed it on MinIO, now they are too lazy to switch over. That also happens. But the real reason that why it became serious for me—I ignored that the public cloud commercialization; I encouraged the community adoption. And it grew to more than a million instances, like across the cloud, like small and large, but when they start talking about paying us serious dollars, then I took it seriously. And then when I start asking them, why would you guys do it, then I got to know the real reason why they wanted to do was they want to be detached from the cloud infrastructure provider.


They want to look at cloud as CPU network and drive as a service. And running their own enterprise IT was more expensive than adopting public cloud, it was productivity for them, reducing the infrastructure, people cost was a lot. It made economic sense.


Corey: Oh, people always cost more the infrastructure itself does.


AB: Exactly right. 70, 80%, like, goes into people, right? And enterprise IT is too slow. They cannot innovate fast, and all of those problems. But what I found was for us, while we actually build the community and customers, if you’re on AWS, if you’re running MinIO on EBS, EBS is three times more expensive than S3.


Corey: Or a single copy of it, too, where if you’re trying to go multi-AZ and you have the replication traffic, and not to mention you have to over-provision it, which is a bit of a different story as well. So, like, it winds up being something on the order of 30 times more expensive, in many cases, to do it right. So, I’m looking at this going, the economics of running this purely by itself in AWS don’t make sense to me—long experience teaches me the next question of, “What am I missing?” Not, “That’s ridiculous and you’re doing it wrong.” There’s clearly something I’m not getting. What am I missing?


AB: I was telling them until we made some changes, right—because we saw a couple of things happen. I was initially like, [unintelligible 00:12:00] does not make 30 copies. It makes, like, 1.4x, 1.6x.


But still, the underlying block storage is not only three times more expensive than S3, it’s also slow. It’s a network storage. Trying to put an object store on top of it, another, like, software-defined SAN, like EBS made no sense to me. Smaller deployments, it’s okay, but you should never scale that on EBS. So, it did not make economic sense. I would never take it seriously because it would never help them grow to scale.


But what changed in recent times? Amazon saw that this was not only a problem for MinIO-type players. Every database out there today, every modern database, even the message queues like Kafka, they all have gone scale-out. And they all depend on local block store and putting a scale-out distributed database, data processing engines on top of EBS would not scale. And Amazon introduced storage optimized instances. Essentially, that reduced to bet—the data infrastructure guy, data engineer, or application developer asking IT, “I want a SuperMicro, or Dell server, or even virtual machines.” That’s too slow, too inefficient.


They can provision these storage machines on demand, and then I can do it through Kubernetes. These two changes, all the public cloud players now adopted Kubernetes as the standard, and they have to stick to the Kubernetes API standard. If they are incompatible, they won’t get adopted. And storage optimized that is local drives, these are machines, like, [I3 EN 00:13:23], like, 24 drives, they have SSDs, and fast network—like, 25-gigabit 200-gigabit type network—availability of these machines, like, what typically would run any database, HDFS cluster, MinIO, all of them, those machines are now available just like any other EC2 instance.


They are efficient. You can actually put MinIO side by side to S3 and still be price competitive. And Amazon wants to—like, just like their retail marketplace, they want to compete and be open. They have enabled it. In that sense, Amazon is actually helping us. And it turned out that now I can help customers build multiple petabyte infrastructure on Amazon and still stay efficient, still stay price competitive.


Corey: I would have said for a long time that if you were to ask me to build out the lingua franca of all the different cloud providers into a common API, the S3 API would be one of them. Now, you are building this out, multi-cloud, you’re in all three of the major cloud marketplaces, and the way that you do that and do those deployments seems like it is the modern multi-cloud API of Kubernetes. When you first started building this, Kubernetes was very early on. What was the evolution of getting there? Or were you one of the first early-adoption customers in a Kubernetes space?


AB: So, when we started, there was no Kubernetes. But we saw the problem was very clear. And there was containers, and then came Docker Compose and Swarm. Then there was Mesos, Cloud Foundry, you name it, right? Like, there was many solutions all the way up to even VMware trying to get into that space.


And what did we do? Early on, I couldn’t choose. I couldn’t—it’s not in our hands, right, who is going to be the winner, so we just simply embrace everybody. It was also tiring that to allow implement native connectors to all of them different orchestration, like Pivotal Cloud Foundry alone, they have their own standard open service broker that’s only popular inside their system. Go outside elsewhere, 
everybody was incompatible.


And outside that, even, Chef Ansible Puppet scripts, too. We just simply embraced everybody until the dust settle down. When it settled down, clearly a declarative model of Kubernetes became easier. Also Kubernetes developers understood the community well. And coming from Borg, I think they understood the right architecture. And also written in Go, unlike Java, right?


It actually matters, these minute new details resonating with the infrastructure community. It took off, and then that helped us immensely. Now, it’s not only Kubernetes is popular, it has become the standard, from VMware to OpenShift to all the public cloud providers, GKS, AKS, EKS, whatever, right—GKE. All of them now are basically Kubernetes standard. It made not only our life easier, it made every other [ISV 00:16:11], other open-source project, everybody now can finally write one code that can be operated portably.


It is a big shift. It is not because we chose; we just watched all this, we were riding along the way. And then because we resonated with the infrastructure community, modern infrastructure is dominated by open-source. We were also the leading open-source object store, and as Kubernetes community adopted us, we were naturally embraced by the community.


Corey: Back when AWS first launched with S3 as its first offering, there were a bunch of folks who were super excited, but object stores didn’t make a lot of sense to them intrinsically, so they looked into this and, “Ah, I can build a file system and users base on top of S3.” And the reaction was, “Holy God don’t do that.” And the way that AWS decided to discourage that behavior is a per request charge, which for most workloads is fine, whatever, but there are some that causes a significant burden. With running something like MinIO in a self-hosted way, suddenly that costing doesn’t exist in the same way. Does that open the door again to so now I can use it as a file system again, in which case that just seems like using the local file system, only with extra steps?


AB: Yeah.


Corey: Do you see patterns that are emerging with customers' use of MinIO that you would not see with the quote-unquote, “Provider’s” quote-unquote, “Native” object storage option, or do the patterns mostly look the same?


AB: Yeah, if you took an application that ran on file and block and brought it over to object storage, that makes sense. But something that is competing with object store or a layer below object store, that is—end of the day that drives our block devices, you have a block interface, right—trying to bring SAN or NAS on top of object store is actually a step backwards. They completely missed the message that Amazon told that if you brought a file system interface on top of object store, you missed the point, that you are now bringing the legacy things that Amazon intentionally removed from the infrastructure. Trying to bring them on top doesn’t make it any better. If you are arguing from a compatibility some legacy applications, sure, but writing a file system on top of object store will never be better than NetApp, EMC, like EMC Isilon, or anything else. Or even GlusterFS, right?


But if you want a file system, I always tell the community, they ask us, “Why don’t you add an FS option and do a multi-protocol system?” I tell them that the whole point of S3 is to remove all those legacy APIs. If I added POSIX, then I’ll be a mediocre object storage and a terrible file system. I would never do that. But why not write a FUSE file system, right? Like, S3Fs is there.


In fact, initially, for legacy compatibility, we wrote MinFS and I had to hide it. We actually archived the repository because immediately people started using it. Even simple things like end of the day, can I use Unix [Coreutils 00:19:03] like [cp, ls 00:19:04], like, all these tools I’m familiar with? If it’s not file system object storage that S3 [CMD 00:19:08] or AWS CLI is, like, to bloatware. And it’s not really Unix-like feeling.


Then what I told them, “I’ll give you a BusyBox like a single static binary, and it will give you all the Unix tools that works for local filesystem as well as object store.” That’s where the [MC tool 00:19:23] came; it gives you all the Unix-like programmability, all the core 
tool that’s object storage compatible, speaks native object store. But if I have to make object store look like a file system so UNIX tools would run, it would not only be inefficient, Unix tools never scaled for this kind of capacity.


So, it would be a bad idea to take step backwards and bring legacy stuff back inside. For some very small case, if there are simple POSIX calls using [ObjectiveFs 00:19:49], S3Fs, and few, for legacy compatibility reasons makes sense, but in general, I would tell the community don’t bring file and block. If you want file and block, leave those on virtual machines and leave that infrastructure in a silo and gradually phase them out.


Corey: This episode is sponsored in part by our friends at Vultr. Spelled V-U-L-T-R because they’re all about helping save money, including on things like, you know, vowels. So, what they do is they are a cloud provider that provides surprisingly high performance cloud compute at a price that—while sure they claim its better than AWS pricing—and when they say that they mean it is less money. Sure, I don’t dispute that but what I find interesting is that it’s predictable. They tell you in advance on a monthly basis what it’s going to going to cost. They have a bunch of advanced networking features. They have nineteen global locations and scale things elastically. Not to be confused with openly, because apparently elastic and open can mean the same thing sometimes. They have had over a million users. Deployments take less that sixty seconds across twelve pre-selected operating systems. Or, if you’re one of those nutters like me, you can bring your own ISO and install basically any operating system you want. Starting with pricing as low as $2.50 a month for Vultr cloud compute they have plans for developers and businesses of all sizes, except maybe Amazon, who stubbornly insists on having something to scale all on their own. Try Vultr today for free by visiting: vultr.com/screaming, and you’ll receive a $100 in credit. Thats v-u-l-t-r.com slash screaming.


Corey: So, my big problem, when I look at what S3 has done is in it’s name because of course, naming is hard. It’s, “Simple Storage Service.” The problem I have is with the word simple because over time, S3 has gotten more and more complex under the hood. It automatically tiers data the way that customers want. And integrated with things like Athena, you can now query it directly, whenever of an object appears, you can wind up automatically firing off Lambda functions and the rest.


And this is increasingly looking a lot less like a place to just dump my unstructured data, and increasingly, a lot like this is sort of a database, in some respects. Now, understand my favorite database is Route 53; I have a long and storied history of misusing services as databases. Is this one of those scenarios, or is there some legitimacy to the idea of turning this into a database?


AB: Actually, there is now S3 Select API that if you’re storing unstructured data like CSV, JSON, Parquet, without downloading even a compressed CSV, you can actually send a SQL query into the system. IN MinIO particularly the S3 Select is [CMD 00:21:16] optimized. We can load, like, every 64k worth of CSV lines into registers and do CMD operations. It’s the fastest SQL filter out there. Now, bringing these kinds of capabilities, we are just a little bit away from a database; should we do database? I would tell definitely no.


The very strength of S3 API is to actually limit all the mutations, right? Particularly if you look at database, they’re dealing with metadata, and querying; the biggest value they bring is indexing the metadata. But if I’m dealing with that, then I’m dealing with really small block lots of mutations, the separation of objects storage should be dealing with persistence and not mutations. Mutations are [AWS 00:21:57] problem. Separation of database work function and persistence function is where object storage got the storage right.


Otherwise, it will, they will make the mistake of doing POSIX-like behavior, and then not only bringing back all those capabilities, doing IOPS intensive workloads across the HTTP, it wouldn’t make sense, right? So, object storage got the API right. But now should it be a database? So, it definitely should not be a database. In fact, I actually hate the idea of Amazon yielding to the file system developers and giving a [file three 00:22:29] hierarchical namespace so they can write nice file managers.


That was a terrible idea. Writing a hierarchical namespace that’s also sorted, now puts tax on how the metadata is indexed and organized. The Amazon should have left the core API very simple and told them to solve these problems outside the object store. Many application developers don’t need. Amazon was trying to satisfy everybody’s need. Saying no to some of these file system-type, file manager-type users, what should have been the right way.


But nevertheless, adding those capabilities, eventually, now you can see, S3 is no longer simple. And we had to keep that compatibility, and I hate that part. I actually don’t mind compatibility, but then doing all the wrong things that Amazon is adding, now I have to add because it’s compatible. I kind of hate that, right?


But now going to a database would be pushing it to the whole new level. Here is the simple reason why that’s a bad idea. The right way to do database—in fact, the database industry is already going in the right direction. Unstructured data, the key-value or graph, different types of data, you cannot possibly solve all that even in a single database. They are trying to be multimodal database; even they are struggling with it.


You can never be a Redis, Cassandra, like, a SQL all-in-one. They tried to say that but in reality, that you will never be better than any one of those focused database solutions out there. Trying to bring that into object store will be a mistake. Instead, let the databases focus on query language implementation and query computation, and leave the persistence to object store. So, object store can still focus on storing your database segments, the table segments, but the index is still in the memory of the database.


Even the index can be snapshotted once in a while to object store, but use objects store for persistence and database for query is the right architecture. And almost all the modern databases now, from Elasticsearch to [unintelligible 00:24:21] to even Kafka, like, message queue. They all have gone that route. Even Microsoft SQL Server, Teradata, Vertica, name it, Splunk, they all have gone object storage route, too. Snowflake itself is a prime example, BigQuery and all of them.


That’s the right way. Databases can never be consolidated. There will be many different kinds of databases. Let them specialize on GraphQL or Graph API, or key-value, or SQL. Let them handle the indexing and persistence, they cannot handle petabytes of data. That [unintelligible 00:24:51] to object store is how the industry is shaping up, and it is going in the right direction.


Corey: One of the ways I learned the most about various services is by talking to customers. Every time I think I’ve seen something, this is amazing. This service is something I completely understand. All I have to do is talk to one more customer. And when I was doing a bill analysis project a couple of years ago, I looked into a customer’s account and saw a bucket with okay, that has 280 billion objects in it—and wait was that billion with a B?


And I asked them, “So, what’s going on over there?” And there’s, “Well, we built our own columnar database on top of S3. This may not have been the best approach.” It’s, “I’m going to stop you there. With no further context, it was not, but please continue.”


It’s the sort of thing that would never have occurred to me to even try, do you tend to see similar—I would say they’re anti-patterns, except somehow they’re made to work—in some of your customer environments, as they are using the service in ways that are very different than ways encouraged or even allowed by the native object store options?


AB: Yeah, when I first started seeing the database-type workloads coming on to MinIO, I was surprised, too. That was exactly my reaction. In fact, they were storing these 256k, sometimes 64k table segments because they need to index it, right, and the table segments were anywhere between 64k to 2MB. And when they started writing table segments, it was more often [IOPS-type 00:26:22] I/O pattern, then a throughput-type pattern. Throughput is an easier problem to solve, and MinIO always saturated these 100-gigabyte NVMe-type drives, they were I/O intensive, throughput optimized.


When I started seeing the database workloads, I had to optimize for small-object workloads, too. We actually did all that because eventually I got convinced the right way to build a database was to actually leave the persistence out of database; they made actually a compelling argument. If historically, I thought metadata and data, data to be very big and coming to object store make sense. Metadata should be stored in a database, and that’s only index page. Take any book, the index pages are only few, database can continue to run adjacent to object store, it’s a clean architecture.


But why would you put database itself on object store? When I saw a transactional database like MySQL, changing the [InnoDB 00:27:14] to [RocksDB 00:27:15], and making changes at that layer to write the SS tables [unintelligible 00:27:19] to MinIO, and then I was like, where do you store the memory, the journal? They said, “That will go to Kafka.” And I was like—I thought that was insane when it started. But it continued to grow and grow.


Nowadays, I see most of the databases have gone to object store, but their argument is, the databases also saw explosive growth in data. And they couldn’t scale the persistence part. That is where they realized that they still got very good at the indexing part that object storage would never give. There is no API to do sophisticated query of the data. You cannot peek inside the data, you can just do streaming read and write.


And that is where the databases were still necessary. But databases were also growing in data. One thing that triggered this was the use case moved from data that was generated by people to now data generated by machines. Machines means applications, all kinds of devices. Now, it’s like between seven billion people to a trillion devices is how the industry is changing. And this led to lots of machine-generated, semi-structured, structured data at giant scale, coming into database. The databases need to handle scale. There was no other way to solve this problem other than leaving the—[unintelligible 00:28:31] if you looking at columnar data, most of them are machine-generated data, where else would you store? If they tried to build their own object storage embedded into the database, it would make database mentally complicated. Let them focus on what they are good at: Indexing and mutations. Pull the data table segments which are immutable, mutate in memory, and then commit them back give the right mix. What you saw what’s the fastest step that happened, we saw that consistently across. Now, it is actually the standard.


Corey: So, you started working on this in 2014, and here we are—what is it—eight years later now, and you’ve just announced a Series B of $100 million dollars on a billion-dollar valuation. So, it turns out this is not just one of those things people are using for test labs; there is significant momentum behind using this. How did you get there from—because everything you’re saying makes an awful lot of sense, but it feels, at least from where I sit, to be a little bit of a niche. It’s a bit of an edge case that is not the common case. Obviously, I missing something because your investors are not the types of sophisticated investors who see something ridiculous and, “Yep. That’s the thing we’re going to go for.” There right more than they’re not.


AB: Yeah. The reason for that was the saw what we were set to do. In fact, these are—if you see the lead investor, Intel, they watched us grow. They came into Series A and they saw, everyday, how we operated and grew. They believed in our message.


And it was actually not about object store, right? Object storage was a means for us to get into the market. When we started, our idea was, ten years from now, what will be a big problem? A lot of times, it’s hard to see the future, but if you zoom out, it’s hidden in plain sight.


These are simple trends. Every major trend pointed to world producing more data. No one would argue with that. If I solved one important problem that everybody is suffering, I won’t go wrong. And when you solve the problem, it’s about building a product with fine craftsmanship, attention to details, connecting with the user, all of that standard stuff.


But I picked object storage as the problem because the industry was fragmented across many different data stores, and I knew that won’t be the case ten years from now. Applications are not going to adopt different APIs across different clouds, S3 to GCS to Azure Blob to HDFS to everything is incompatible. I saw that if I built a data store for persistence, industry will consolidate around S3 API. Amazon S3, when we started, it looked like they were the giant, there was only one cloud industry, it believed mono-cloud. Almost everyone was talking to me like AWS will be the world’s data center.


I certainly see that possibility, Amazon is capable of doing it, but my bet was the other way, that AWS S3 will be one of many solutions, but not—if it’s all incompatible, it’s not going to work, industry will consolidate. Our bet was, if world is producing so much data, if you build an object store that is S3 compatible, but ended up as the leading data store of the world and owned the application ecosystem, you cannot go wrong. We kept our heads low and focused on the first six years on massive adoption, build the ecosystem to a scale where we can say now our ecosystem is equal or larger than Amazon, then we are in business. We didn’t focus on commercialization; we focused on convincing the industry that this is the right technology for them to use. Once they are convinced, once you solve business problems, making money is not hard because they are already sold, they are in love with the product, then convincing them to pay is not a big deal because data is so critical, central part of their business.


We didn’t worry about commercialization, we worried about adoption. And once we got the adoption, now customers are coming to us and they’re like, “I don’t want open-source license violation. I don’t want data breach or data loss.” They are trying to sell to me, and it’s an easy relationship game. And it’s about long-term partnership with customers.


And so the business started growing, accelerating. That was the reason that now is the time to fill up the gas tank and investors were quite excited about the commercial traction as well. And all the intangible, right, how big we grew in the last few years.


Corey: It really is an interesting segment, that has always been something that I’ve mostly ignored, like, “Oh, you want to run your own? Okay, great.” I get it; some people want to cosplay as cloud providers themselves. Awesome. There’s clearly a lot more to it than that, and I’m really interested to see what the future holds for you folks.


AB: Yeah, I’m excited. I think end of the day, if I solve real problems, every organization is moving from compute technology-centric to data-centric, and they’re all looking at data warehouse, data lake, and whatever name they give data infrastructure. Data is now the centerpiece. Software is a commodity. That’s how they are looking at it. And it is translating to each of these large organizations—actually, even the mid, even startups nowadays have petabytes of data—and I see a huge potential here. The timing is perfect for us.


Corey: I’m really excited to see this continue to grow. And I want to thank you for taking so much time to speak with me today. If people want to learn more, where can they find you?


AB: I’m always on the community, right. Twitter and, like, I think the Slack channel, it’s quite easy to reach out to me. LinkedIn. I’m always excited to talk to our users or community.


Corey: And we will of course put links to this in the [show notes 00:33:58]. Thank you so much for your time. I really appreciate it.


AB: Again, wonderful to be here, Corey.


Corey: Anand Babu Periasamy, CEO and co-founder of MinIO. I’m Cloud Economist Corey Quinn, and this is Screaming in the Cloud. If you’ve enjoyed this podcast, please leave a five-star review on your podcast platform of choice, whereas if you’ve hated this podcast, please leave a five-star review on your podcast platform of choice, along with what starts out as an angry comment but eventually turns into you, in your position on the S3 product team, writing a thank you note to MinIO for helping validate your market.


Corey: If your AWS bill keeps rising and your blood pressure is doing the same, then you need The Duckbill Group. We help companies fix their AWS bill by making it smaller and less horrifying. The Duckbill Group works for you, not AWS. We tailor recommendations to your business and we get to the point. Visit duckbillgroup.com to get started.


Announcer: This has been a HumblePod production. Stay humble.
Newsletter Footer

Get the Newsletter

Reach over 30,000 discerning engineers, managers, enthusiasts who actually care about the state of Amazon’s cloud ecosystems.

This field is for validation purposes and should be left unchanged.
Sponsor Icon Footer

Sponsor an Episode

Get your message in front of people who care enough to keep current about the cloud phenomenon and its business impacts.