>> Paige: Hello, everybody and thank you for joining us today on the Virtual Vertica BDC 2020. Today's breakout session is entitled A Technical Overview of the Vertica Architecture. I'm Paige Roberts, Open Source Relations Manager at Vertica and I'll be your host for this webinar. Now joining me is Ryan Role-kuh? Did I say that right? (laughs) He's a Vertica Senior Software Engineer. >> Ryan: So it's Roelke. (laughs) >> Paige: Roelke, okay, I got it, all right. Ryan Roelke. And before we begin, I want to be sure and encourage you guys to submit your questions or your comments during the virtual session while Ryan is talking as you think of them as you go along. You don't have to wait to the end, just type in your question or your comment in the question box below the slides and click submit. There'll be a Q and A at the end of the presentation and we'll answer as many questions as we're able to during that time. Any questions that we don't address, we'll do our best to get back to you offline. Now, alternatively, you can visit the Vertica forums to post your question there after the session as well. Our engineering team is planning to join the forums to keep the conversation going, so you can have a chat afterwards with the engineer, just like any other conference. Now also, you can maximize your screen by clicking the double arrow button in the lower right corner of the slides and before you ask, yes, this virtual session is being recorded and it will be available to view on demand this week. We'll send you a notification as soon as it's ready. Now, let's get started. Over to you, Ryan. >> Ryan: Thanks, Paige. Good afternoon, everybody. My name is Ryan and I'm a Senior Software Engineer on Vertica's Development Team. I primarily work on improving Vertica's query execution engine, so usually in the space of making things faster. Today, I'm here to talk about something that's more general than that, so we're going to go through a technical overview of the Vertica architecture. So the intent of this talk, essentially, is to just explain some of the basic aspects of how Vertica works and what makes it such a great database software and to explain what makes a query execute so fast in Vertica, we'll provide some background to explain why other databases don't keep up. And we'll use that as a starting point to discuss an academic database that paved the way for Vertica. And then we'll explain how Vertica design builds upon that academic database to be the great software that it is today. I want to start by sharing somebody's approximation of an internet minute at some point in 2019. All of the data on this slide is generated by thousands or even millions of users and that's a huge amount of activity. Most of the applications depicted here are backed by one or more databases. Most of this activity will eventually result in changes to those databases. For the most part, we can categorize the way these databases are used into one of two paradigms. First up, we have online transaction processing or OLTP. OLTP workloads usually operate on single entries in a database, so an update to a retail inventory or a change in a bank account balance are both great examples of OLTP operations. Updates to these data sets must be visible immediately and there could be many transactions occurring concurrently from many different users. OLTP queries are usually key value queries. The key uniquely identifies the single entry in a database for reading or writing. Early databases and applications were probably designed for OLTP workloads. This example on the slide is typical of an OLTP workload. We have a table, accounts, such as for a bank, which tracks information for each of the bank's clients. An update query, like the one depicted here, might be run whenever a user deposits $10 into their bank account. Our second category is online analytical processing or OLAP which is more about using your data for decision making. If you have a hardware device which periodically records how it's doing, you could analyze trends of all your devices over time to observe what data patterns are likely to lead to failure or if you're Google, you might log user search activity to identify which links helped your users find the answer. Analytical processing has always been around but with the advent of the internet, it happened at scales that were unimaginable, even just 20 years ago. This SQL example is something you might see in an OLAP workload. We have a table, searches, logging user activity. We will eventually see one row in this table for each query submitted by users. If we want to find out what time of day our users are most active, then we could write a query like this one on the slide which counts the number of unique users running searches for each hour of the day. So now let's rewind to 2005. We don't have a picture of an internet minute in 2005, we don't have the data for that. We also don't have the data for a lot of other things. The term Big Data is not quite yet on anyone's radar and The Cloud is also not quite there or it's just starting to be. So if you have a database serving your application, it's probably optimized for OLTP workloads. OLAP workloads just aren't mainstream yet and database engineers probably don't have them in mind. So let's innovate. It's still 2005 and we want to try something new with our database. Let's take a look at what happens when we do run an analytic workload in 2005. Let's use as a motivating example a table of stock prices over time. In our table, the symbol column identifies the stock that was traded, the price column identifies the new price and the timestamp column indicates when the price changed. We have several other columns which, we should know that they're there, but we're not going to use them in any example queries. This table is designed for analytic queries. We're probably not going to make any updates or look at individual rows since we're logging historical data and want to analyze changes in stock price over time. Our database system is built to serve OLTP use cases, so it's probably going to store the table on disk in a single file like this one. Notice that each row contains all of the columns of our data in row major order. There's probably an index somewhere in the memory of the system which will help us to point lookups. Maybe our system expects that we will use the stock symbol and the trade time as lookup keys. So an index will provide quick lookups for those columns to the position of the whole row in the file. If we did have an update to a single row, then this representation would work great. We would seek to the row that we're interested in, finding it would probably be very fast using the in-memory index. And then we would update the file in place with our new value. On the other hand, if we ran an analytic query like we want to, the data access pattern is very different. The index is not helpful because we're looking up a whole range of rows, not just a single row. As a result, the only way to find the rows that we actually need for this query is to scan the entire file. We're going to end up scanning a lot of data that we don't need and that won't just be the rows that we don't need, there's many other columns in this table. Many information about who made the transaction, and we'll also be scanning through those columns for every single row in this table. That could be a very serious problem once we consider the scale of this file. Stocks change a lot, we probably have thousands or millions or maybe even billions of rows that are going to be stored in this file and we're going to scan all of these extra columns for every single row. If we tried out our stocks use case behind the desk for the Fortune 500 company, then we're probably going to be pretty disappointed. Our queries will eventually finish, but it might take so long that we don't even care about the answer anymore by the time that they do. Our database is not built for the task we want to use it for. Around the same time, a team of researchers in the North East have become aware of this problem and they decided to dedicate their time and research to it. These researchers weren't just anybody. The fruits of their labor, which we now like to call the C-Store Paper, was published by eventual Turing Award winner, Mike Stonebraker, along with several other researchers from elite universities. This paper presents the design of a read-optimized relational DBMS that contrasts sharply with most current systems, which are write-optimized. That sounds exactly like what we want for our stocks use case. Reasoning about what makes our queries executions so slow brought our researchers to the Memory Hierarchy, which essentially is a visualization of the relative speeds of different parts of a computer. At the top of the hierarchy, we have the fastest data units, which are, of course, also the most expensive to produce. As we move down the hierarchy, components get slower but also much cheaper and thus you can have more of them. Our OLTP databases data is stored in a file on the hard disk. We scanned the entirety of this file, even though we didn't need most of the data and now it turns out, that is just about the slowest thing that our query could possibly be doing by over two orders of magnitude. It should be clear, based on that, that the best thing we can do to optimize our query's execution is to avoid reading unnecessary data from the disk and that's what the C-Store researchers decided to look at. The key innovation of the C-Store paper does exactly that. Instead of storing data in a row major order, in a large file on disk, they transposed the data and stored each column in its own file. Now, if we run the same select query, we read only the relevant columns. The unnamed columns don't factor into the table scan at all since we don't even open the files. Zooming out to an internet scale sized data set, we can appreciate the savings here a lot more. But we still have to read a lot of data that we don't need to answer this particular query. Remember, we had two predicates, one on the symbol column and one on the timestamp column. Our query is only interested in AAPL stock, but we're still reading rows for all of the other stocks. So what can we do to optimize our disk read even more? Let's first partition our data set into different files based on the timestamp date. This means that we will keep separate files for each date. When we query the stocks table, the database knows all of the files we have to open. If we have a simple predicate on the timestamp column, as our sample query does, then the database can use it to figure out which files we don't have to look at at all. So now all of our disk reads that we have to do to answer our query will produce rows that pass the timestamp predicate. This eliminates a lot of wasteful disk reads. But not all of them. We do have another predicate on the symbol column where symbol equals AAPL. We'd like to avoid disk reads of rows that don't satisfy that predicate either. And we can avoid those disk reads by clustering all the rows that match the symbol predicate together. If all of the AAPL rows are adjacent, then as soon as we see something different, we can stop reading the file. We won't see any more rows that can pass the predicate. Then we can use the positions of the rows we did find to identify which pieces of the other columns we need to read. One technique that we can use to cluster the rows is sorting. So we'll use the symbol column as a sort key for all of the columns. And that way we can reconstruct a whole row by seeking to the same row position in each file. It turns out, having sorted all of the rows, we can do a bit more. We don't have any more wasted disk reads but we can still be more efficient with how we're using the disk. We've clustered all of the rows with the same symbol together so we don't really need to bother repeating the symbol so many times in the same file. Let's just write the value once and say how many rows we have. This one length encoding technique can compress large numbers of rows into a small amount of space. In this example, we do de-duplicate just a few rows but you can imagine de-duplicating many thousands of rows instead. This encoding is great for reducing the amounts of disk we need to read at query time, but it also has the additional benefit of reducing the total size of our stored data. Now our query requires substantially fewer disk reads than it did when we started. Let's recap what the C-Store paper did to achieve that. First, we transposed our data to store each column in its own file. Now, queries only have to read the columns used in the query. Second, we partitioned the data into multiple file sets so that all rows in a file have the same value for the partition column. Now, a predicate on the partition column can skip non-matching file sets entirely. Third, we selected a column of our data to use as a sort key. Now rows with the same value for that column are clustered together, which allows our query to stop reading data once it finds non-matching rows. Finally, sorting the data this way enables high compression ratios, using one length encoding which minimizes the size of the data stored on the disk. The C-Store system combined each of these innovative ideas to produce an academically significant result. And if you used it behind the desk of a Fortune 500 company in 2005, you probably would've been pretty pleased. But it's not 2005 anymore and the requirements of a modern database system are much stricter. So let's take a look at how C-Store fairs in 2020. First of all, we have designed the storage layer of our database to optimize a single query in a single application. Our design optimizes the heck out of that query and probably some similar ones but if we want to do anything else with our data, we might be in a bit of trouble. What if we just decide we want to ask a different question? For example, in our stock example, what if we want to plot all the trade made by a single user over a large window of time? How do our optimizations for the previous query measure up here? Well, our data's partitioned on the trade date, that could still be useful, depending on our new query. If we want to look at a trader's activity over a long period of time, we would have to open a lot of files. But if we're still interested in just a day's worth of data, then this optimization is still an optimization. Within each file, our data is ordered on the stock symbol. That's probably not too useful anymore, the rows for a single trader aren't going to be clustered together so we will have to scan all of the rows in order to figure out which ones match. You could imagine a worse design but as it becomes crucial to optimize this new type of query, then we might have to go as far as reconfiguring the whole database. The next problem of one of scale. One server is probably not good enough to serve a database in 2020. C-Store, as described, runs on a single server and stores lots of files. What if the data overwhelms this small system? We could imagine exhausting the file system's inodes limit with lots of small files due to our partitioning scheme. Or we could imagine something simpler, just filling up the disk with huge volumes of data. But there's an even simpler problem than that. What if something goes wrong and C-Store crashes? Then our data is no longer available to us until the single server is brought back up. A third concern, another one of scalability, is that one deployment does not really suit all possible things and use cases we could imagine. We haven't really said anything about being flexible. A contemporary database system has to integrate with many other applications, which might themselves have pretty restricted deployment options. Or the demands imposed by our workloads have changed and the setup you had before doesn't suit what you need now. C-Store doesn't do anything to address these concerns. What the C-Store paper did do was lead very quickly to the founding of Vertica. Vertica's architecture and design are essentially all about bringing the C-Store designs into an enterprise software system. The C-Store paper was just an academic exercise so it didn't really need to address any of the hard problems that we just talked about. But Vertica, the first commercial database built upon the ideas of the C-Store paper would definitely have to. This brings us back to the present to look at how an analytic query runs in 2020 on the Vertica Analytic Database. Vertica takes the key idea from the paper, can we significantly improve query performance by changing the way our data is stored and give its users the tools to customize their storage layer in order to heavily optimize really important or commonly wrong queries. On top of that, Vertica is a distributed system which allows it to scale up to internet-sized data sets, as well as have better reliability and uptime. We'll now take a brief look at what Vertica does to address the three inadequacies of the C-Store system that we mentioned. To avoid locking into a single database design, Vertica provides tools for the database user to customize the way their data is stored. To address the shortcomings of a single node system, Vertica coordinates processing among multiple nodes. To acknowledge the large variety of desirable deployments, Vertica does not require any specialized hardware and has many features which smoothly integrate it with a Cloud computing environment. First, we'll look at the database design problem. We're a SQL database, so our users are writing SQL and describing their data in SQL way, the Create Table statement. Create Table is a logical description of what your data looks like but it doesn't specify the way that it has to be stored, For a single Create Table, we could imagine a lot of different storage layouts. Vertica adds some extensions to SQL so that users can go even further than Create Table and describe the way that they want the data to be stored. Using terminology from the C-Store paper, we provide the Create Projection statement. Create Projection specifies how table data should be laid out, including column encoding and sort order. A table can have multiple projections, each of which could be ordered on different columns. When you query a table, Vertica will answer the query using the projection which it determines to be the best match. Referring back to our stock example, here's a sample Create Table and Create Projection statement. Let's focus on our heavily optimized example query, which had predicates on the stock symbol and date. We specify that the table data is to be partitioned by date. The Create Projection Statement here is excellent for this query. We specify using the order by clause that the data should be ordered according to our predicates. We'll use the timestamp as a secondary sort key. Each projection stores a copy of the table data. If you don't expect to need a particular column in a projection, then you can leave it out. Our average price query didn't care about who did the trading, so maybe our projection design for this query can leave the trader column out entirely. If the question we want to ask ever does change, maybe we already have a suitable projection, but if we don't, then we can create another one. This example shows another projection which would be much better at identifying trends of traders, rather than identifying trends for a particular stock. Next, let's take a look at our second problem, that one, or excuse me, so how should you decide what design is best for your queries? Well, you could spend a lot of time figuring it out on your own, or you could use Vertica's Database Designer tool which will help you by automatically analyzing your queries and spitting out a design which it thinks is going to work really well. If you want to learn more about the Database Designer Tool, then you should attend the session Vertica Database Designer- Today and Tomorrow which will tell you a lot about what the Database Designer does and some recent improvements that we have made. Okay, now we'll move to our next problem. (laughs) The challenge that one server does not fit all. In 2020, we have several orders of magnitude more data than we had in 2005. And you need a lot more hardware to crunch it. It's not tractable to keep multiple petabytes of data in a system with a single server. So Vertica doesn't try. Vertica is a distributed system so will deploy multiple severs which work together to maintain such a high data volume. In a traditional Vertica deployment, each node keeps some of the data in its own locally-attached storage. Data is replicated so that there is a redundant copy somewhere else in the system. If any one node goes down, then the data that it served is still available on a different node. We'll also have it so that in the system, there's no special node with extra duties. All nodes are created equal. This ensures that there is no single point of failure. Rather than replicate all of your data, Vertica divvies it up amongst all of the nodes in your system. We call this segmentation. The way data is segmented is another parameter of storage customization and it can definitely have an impact upon query performance. A common way to segment data is by using a hash expression, which essentially randomizes the node that a row of data belongs to. But with a guarantee that the same data will always end up in the same place. Describing the way data is segmented is another part of the Create Projection Statement, as seen in this example. Here we segment on the hash of the symbol column so all rows with the same symbol will end up on the same node. For each row that we load into the system, we'll apply our segmentation expression. The result determines which segment the row belongs to and then we'll send the row to each node which holds the copy of that segment. In this example, our projection is marked KSAFE 1, so we will keep one redundant copy of each segment. When we load a row, we might find that its segment had copied on Node One and Node Three, so we'll send a copy of the row to each of those nodes. If Node One is temporarily disconnected from the network, then Node Three can serve the other copy of the segment so that the whole system remains available. The last challenge we brought up from the C-Store design was that one deployment does not fit all. Vertica's cluster design neatly addressed many of our concerns here. Our use of segmentation to distribute data means that a Vertica system can scale to any size of deployment. And since we lack any special hardware or nodes with special purposes, Vertica servers can run anywhere, on premise or in the Cloud. But let's suppose you need to scale out your cluster to rise to the demands of a higher workload. Suppose you want to add another node. This changes the division of the segmentation space. We'll have to re-segment every row in the database to find its new home and then we'll have to move around any data that belongs to a different segment. This is a very expensive operation, not something you want to be doing all that often. Traditional Vertica doesn't solve that problem especially well, but Vertica Eon Mode definitely does. Vertica's Eon Mode is a large set of features which are designed with a Cloud computing environment in mind. One feature of this design is elastic throughput scaling, which is the idea that you can smoothly change your cluster size without having to pay the expenses of shuffling your entire database. Vertica Eon Mode had an entire session dedicated to it this morning. I won't say any more about it here, but maybe you already attended that session or if you haven't, then I definitely encourage you to listen to the recording. If you'd like to learn more about the Vertica architecture, then you'll find on this slide links to several of the academic conference publications. These four papers here, as well as Vertica Seven Years Later paper which describes some of the Vertica designs seven years after the founding and also a paper about the innovations of Eon Mode and of course, the Vertica documentation is an excellent resource for learning more about what's going on in a Vertica system. I hope you enjoyed learning about the Vertica architecture. I would be very happy to take all of your questions now. Thank you for attending this session.

>> Paige: Hello everybody and thank you for joining us today for the virtual Vertica BDC 2020. Today's breakout session is entitled Vertica in Eon Mode past, present and future. I'm Paige Roberts, open source relations manager at Vertica and I'll be your host for this session. Joining me is Vertica engineer, Yuanzhe Bei and Vertica Product Manager, David Sprogis. Before we begin, I encourage you to submit questions or comments during the virtual session. You don't have to wait till the end. Just type your question or comment as you think of it in the question box, below the slides and click Submit. Q&A session at the end of the presentation. We'll answer as many of your questions as we're able to during that time, and any questions that we don't address, we'll do our best to answer offline. If you wish after the presentation, you can visit the Vertica forums to post your questions there and our engineering team is planning to join the forums to keep the conversation going, just like a Dev Lounge at a normal in person, BDC. So, as a reminder, you can maximize your screen by clicking the double arrow button in the lower right corner of the slides, if you want to see them bigger. And yes, before you ask, this virtual session is being recorded and will be available to view on demand this week. We are supposed to send you a notification as soon as it's ready. All right, let's get started. Over to you, Dave. >> David: Thanks, Paige. Hey, everybody. Let's start with a timeline of the life of Eon Mode. About two years ago, a little bit less than two years ago, we introduced Eon Mode on AWS. Pretty specifically for the purpose of rapid scaling to meet the cloud economics promise. It wasn't long after that we realized that workload isolation, a byproduct of the architecture was very important to our users and going to the third tick, you can see that the importance of that workload isolation was manifest in Eon Mode being made available on-premise using Pure Storage FlashBlade. Moving to the fourth tick mark, we took steps to improve workload isolation, with a new type of subcluster which Yuanzhe will go through and to the fifth tick mark, the introduction of secondary subclusters for faster scaling and other improvements which we will cover in the slides to come. Getting started with, why we created Eon Mode in the first place. Let's imagine that your database is this pie, the pecan pie and we're loading pecan data in through the ETL cutting board in the upper left hand corner. We have a couple of free floating pecans, which we might imagine to be data supporting external tables. As you know, the Vertica has a query engine capability as well which we call external tables. And so if we imagine this pie, we want to serve it with a number of servers. Well, let's say we wanted to serve it with three servers, three nodes, we would need to slice that pie into three segments and we would serve each one of those segments from one of our nodes. Now because the data is important to us and we don't want to lose it, we're going to be saving that data on some kind of raid storage or redundant storage. In case one of the drives goes bad, the data remains available because of the durability of raid. Imagine also, that we care about the availability of the overall database. Imagine that a node goes down, perhaps the second node goes down, we still want to be able to query our data and through nodes one and three, we still have all three shards covered and we can do this because of buddy projections. Each neighbor, each nodes neighbor contains a copy of the data from the node next to it. And so in this case, node one is sharing its segment with node two. So node two can cover node one, node three can cover node two and node one back to node three. Adding a little bit more complexity, we might store the data in different copies, each copy sorted for a different kind of query. We call this projections in Vertica and for each projection, we have another copy of the data sorted differently. Now it gets complex. What happens when we want to add a node? Well, if we wanted to add a fourth node here, what we would have to do, is figure out how to re-slice all of the data in all of the copies that we have. In effect, what we want to do is take our three slices and slice it into four, which means taking a portion of each of our existing thirds and re-segmenting into quarters. Now that looks simple in the graphic here, but when it comes to moving data around, it becomes quite complex because for each copy of each segment we need to replace it and move that data on to the new node. What's more, the fourth node can't have a copy of itself that would be problematic in case it went down. Instead, what we need is we need that buddy to be sitting on another node, a neighboring node. So we need to re-orient the buddies as well. All of this takes a lot of time, it can take 12, 24 or even 36 hours in a period when you do not want your database under high demand. In fact, you may want to stop loading data altogether in order to speed it up. This is a planned event and your applications should probably be down during this period, which makes it difficult. With the advent of cloud computing, we saw that services were coming up and down faster and we determined to re-architect Vertica in a way to accommodate that rapid scaling. Let's see how we did it. So let's start with four nodes now and we've got our four nodes database. Let's add communal storage and move each of the segments of data into communal storage. Now that's the separation that we're talking about. What happens if we run queries against it? Well, it turns out that the communal storage is not necessarily performing and so the IO would be slow, which would make the overall queries slow. In order to compensate for the low performance of communal storage, we need to add back local storage, now it doesn't have to be raid because this is just an ephemeral copy but with the data files, local to the node, the queries will run much faster. In AWS, communal storage really does mean an S3 bucket and here's a simplified version of the diagram. Now, do we need to store all of the data from the segment in the depot? The answer is no and the graphics inside the bucket has changed to reflect that. It looks more like a bullseye, showing just a segment of the data being copied to the cache or to the depot, as we call it on each one of the nodes. How much data do you store on the node? Well, it would be the active data set, the last 30 days, the last 30 minutes or the last. Whatever period of time you're working with. The active working set is the hot data and that's how large you want to size your depot. By architecting this way, when you scale up, you're not re-segmenting the database. What you're doing, is you're adding more compute and more subscriptions to the existing shards of the existing database. So in this case, we've added a complete set of four nodes. So we've doubled our capacity and we've doubled our subscriptions, which means that now, the two nodes can serve the yellow shard, two nodes can serve the red shard and so on. In this way, we're able to run twice as many queries in the same amount of time. So you're doubling the concurrency. How high can you scale? Well, can you scale to 3X, 5X? We tested this in the graphics on the right, which shows concurrent users in the X axis by the number of queries executed in a minute along the Y axis. We've grouped execution in runs of 10 users, 30 users, 50, 70 up to 150 users. Now focusing on any one of these groups, particularly up around 150. You can see through the three bars, starting with the bright purple bar, three nodes and three segments. That as you add nodes to the middle purple bar, six nodes and three segments, you've almost doubled your throughput up to the dark purple bar which is nine nodes and three segments and our tests show that you can go to 5X with pretty linear performance increase. Beyond that, you do continue to get an increase in performance but your incremental performance begins to fall off. Eon architecture does something else for us and that is it provides high availability because each of the nodes can be thought of as ephemeral and in fact, each node has a buddy subscription in a way similar to the prior architecture. So if we lose node four, we're losing the node responsible for the red shard and now node one has to pick up responsibility for the red shard while that node is down. When a query comes in, and let's say it comes into one and one is the initiator then one will look for participants, it'll find a blue shard and a green shard but when it's looking for the red, it finds itself and so the node number one will be doing double duty. This means that your performance will be cut in half approximately, for the query. This is acceptable until you are able to restore the node. Once you restore it and once the depot becomes rehydrated, then your performance goes back to normal. So this is a much simpler way to recover nodes in the event of node failure. By comparison, Enterprise Mode the older architecture. When we lose the fourth node, node one takes over responsibility for the first shard and the yellow shard and the red shard. But it also is responsible for rehydrating the entire data segment of the red shard to node four, this can be very time consuming and imposes even more stress on the first node. So performance will go down even further. Eon Mode has another feature and that is you can scale down completely to zero. We call this hibernation, you shut down your database and your database will maintain full consistency in a rest state in your S3 bucket and then when you need access to your database again, you simply recreate your cluster and revive your database and you can access your database once again. That concludes the rapid scaling portion of, why we created Eon Mode. To take us through workload isolation is Yuanzhe Bei, Yuanzhe. >> Yuanzhe: Thanks Dave, for presenting how Eon works in general. In the next section, I will show you another important capability of Vertica Eon Mode, the workload isolation. Dave used a pecan pie as an example of database. Now let's say it's time for the main course. Does anyone still have a problem with food touching on their plates. Parents know that it's a common problem for kids. Well, we have a similar problem in database as well. So there could be multiple different workloads accessing your database at the same time. Say you have ETL jobs running regularly. While at the same time, there are dashboards running short queries against your data. You may also have the end of month report running and their can be ad hoc data scientists, connect to the database and do whatever the data analysis they want to do and so on. How to make these mixed workload requests not interfere with each other is a real challenge for many DBAs. Vertica Eon Mode provides you the solution. I'm very excited here to introduce to you to the important concept in Eon Mode called subclusters. In Eon Mode, nodes they belong to the predefined subclusters rather than the whole cluster. DBAs can define different subcluster for different kinds of workloads and it redirects those workloads to the specific subclusters. For example, you can have an ETL subcluster, dashboard subcluster, report subcluster and the analytic machine learning subcluster. Vertica Eon subcluster is designed to achieve the three main goals. First of all, strong workload isolation. That means any operation in one subcluster should not affect or be affected by other subclusters. For example, say the subcluster running the report is quite overloaded and already there can be, the data scienctists running crazy analytic jobs, machine learning jobs on the analytics subcluster and making it very slow, even stuck or crash or whatever. In such scenario, your ETL and dashboards subcluster should not be or at least very minimum be impacted by this crisis and which means your ETL job which should not lag behind and dashboard should respond timely. We have done a lot of improvements as of 10.0 release and will continue to deliver improvements in this category. Secondly, fully customized subcluster settings. That means any subcluster can be set up and tuned for very different workloads without affecting other subclusters. Users should be able to tune up, tune down, certain parameters based on the actual needs of the individual subcluster workload requirements. As of today, Vertica already supports few settings that can be done at the subcluster level for example, the depot pinning policy and then we will continue extending more that is like resource pools (mumbles) in the near future. Lastly, Vertica subclusters should be easy to operate and cost efficient. What it means is that the subcluster should be able to turn on, turn off, add or remove or should be available for use according to rapid changing workloads. Let's say in this case, you want to spin up more dashboard subclusters because we need higher scores report, we can do that. You might need to run several report subclusters because you might want to run multiple reports at the same time. While on the other hand, you can shut down your analytic machine learning subcluster because no data scientists need to use it at this moment. So we made automate a lot of change, the improvements in this category, which I'll explain in detail later and one of the ultimate goal is to support auto scaling To sum up, what we really want to deliver for subcluster is very simple. You just need to remember that accessing subclusters should be just like accessing individual clusters. Well, these subclusters do share the same catalog. So you don't have to work out the stale data and don't need to worry about data synchronization. That'd be a nice goal, Vertica upcoming 10.0 release is certainly a milestone towards that goal, which will deliver a large part of the capability in this direction and then we will continue to improve it after 10.0 release. In the next couple of slides, I will highlight some issues about workload isolation in the initial Eon release and show you how we resolve these issues. First issue when we initially released our first or so called subcluster mode, it was implemented using fault groups. Well, fault groups and the subcluster have something in common. Yes, they are both defined as a set of nodes. However, they are very different in all the other ways. So, that was very confusing in the first place, when we implement this. As of 9.3.0 version, we decided to detach subcluster definition from the fault groups, which enabled us to further extend the capability of subclusters. Fault groups in the pre 9.3.0 versions will be converted into subclusters during the upgrade and this was a very important step that enabled us to provide all the amazing, following improvements on subclusters. The second issue in the past was that it's hard to control the execution groups for different types of workloads. There are two types of problems here and I will use some example to explain. The first issue is about control group size. There you allocate six nodes for your dashboard subcluster and what you really want is on the left, the three pairs of nodes as three execution groups, and each pair of nodes will need to subscribe to all the four shards. However, that's not really what you get. What you really get is there on the right side that the first four nodes subscribed to one shard each and the rest two nodes subscribed to two dangling shards. So you won't really get three execusion groups but instead only get one and two extra nodes have no value at all. The solution is to use subclusters. So instead of having a subcluster with six nodes, you can split it up into three smaller ones. Each subcluster will guarantee to subscribe to all the shards and you can further handle this three subcluster using load balancer across them. In this way you achieve the three real exclusion groups. The second issue is that the session participation is non-deterministic. Any session will just pick four random nodes from the subcluster as long as this covers one shard each. In other words, you don't really know which set of nodes will make up your execution group. What's the problem? So in this case, the fourth node will be doubled booked by two concurrent sessions. And you can imagine that the resource usage will be imbalanced and both queries performance will suffer. What is even worse is that these queries of the two concurrent sessions target different table They will cause the issue, that depot efficiency will be reduced, because both session will try to fetch the files on to two tables into the same depot and if your depot is not large enough, they will evict each other, which will be very bad. To solve this the same way, you can solve this by declaring subclusters, in this case, two subclusters and a load balancer group across them. The reason it solved the problem is because the session participation would not go across the boundary. So there won't be a case that any node is double booked and in terms of the depot and if you use the subcluster and avoid using a load balancer group, and carefully send the first workload to the first subcluster and the second to the second subcluster and then the result is that depot isolation is achieved. The first subcluster will maintain the data files for the first query and you don't need to worry about the file being evicted by the second kind of session. Here comes the next issue, it's the scaling down. In the old way of defining subclusters, you may have several execution groups in the subcluster. You want to shut it down, one or two execution groups to save cost. Well, here comes the pain, because you don't know which nodes may be used by which session at any point, it is hard to find the right timing to hit the shutdown button of any of the instances. And if you do and get unlucky, say in this case, you pull the first four nodes, one of the session will fail because it's participating in the node two and node four at that point. User of that session will notice because their query fails and we know that for many business this is critical problem and not acceptable. Again, with subclusters this problem is resolved. Same reason, session cannot go across the subcluster boundary. So all you need to do is just first prevent query sent to the first subcluster and then you can shut down the instances in that subcluster. You are guaranteed to not break any running sessions. Now, you're happy and you want to shut down more subclusters then you hit the issue four, the whole cluster will go down, why? Because the cluster loses quorum. As a distributed system, you need to have at least more than half of a node to be up in order to commit and keep the cluster up. This is to prevent the catalog diversion from happening, which is important. But do you still want to shut down those nodes? Because what's the point of keeping those nodes up and if you are not using them and let them cost you money right. So Vertica has a solution, you can define a subcluster as secondary to allow them to shut down without worrying about quorum. In this case, you can define the first three subclusters as secondary and the fourth one as primary. By doing so, this secondary subclusters will not be counted towards the quorum because we changed the rule. Now instead of requiring more than half of node to be up, it only require more than half of the primary node to be up. Now you can shut down your second subcluster and even shut down your third subcluster as well and keep the remaining primary subcluster to be still running healthily. There are actually more benefits by defining secondary subcluster in addition to the quorum concern, because the secondary subclusters no longer have the voting power, they don't need to persist catalog anymore. This means those nodes are faster to deploy, and can be dropped and re-added. Without the worry about the catalog persistency. For the most the subcluster that only need to read only query, it's the best practice to define them as secondary. The commit will be faster on this secondary subcluster as well, so running this query on the secondary subcluster will have less spikes. Primary subcluster as usual handle everything is responsible for consistency, the background tasks will be running. So DBAs should make sure that the primary subcluster is stable and assume is running all the time. Of course, you need to at least one primary subcluster in your database. Now with the secondary subcluster, user can start and stop as they need, which is very convenient and this further brings up another issue is that if there's an ETL transaction running and in the middle, a subcluster starting and it become up. In older versions, there is no catalog resync mechanism to keep the new subcluster up to date. So Vertica rolls back to ETL session to keep the data consistency. This is actually quite disruptive because real world ETL workloads can sometimes take hours and rolling back at the end means, a large waste of resources. We resolved this issue in 9.3.1 version by introducing a catalog resync mechanism when such situation happens. ETL transactions will not roll back anymore, but instead will take some time to resync the catalog and commit and the problem is resolved. And last issue I would like to talk about is the subscription. Especially for large subcluster when you start it, the startup time is quite long, because the subscription commit used to be serialized. In one of the in our internal testing with large catalogs committing a subscription, you can imagine it takes five minutes. Secondary subcluster is better, because it doesn't need to persist the catalog during the commit but still take about two seconds to commit. So what's the problem here? Let's do the math and look at this chart. The X axis is the time in the minutes and the Y axis is the number of nodes to be subscribed. The dark blues represents your primary subcluster and light blue represents the secondary subcluster. Let's say the subcluster have 16 nodes in total and if you start a secondary subcluster, it will spend about 30 seconds in total, because the 2 seconds times 16 is 32. It's not actually that long time. but if you imagine that starting secondary subcluster, you expect it to be super fast to react to the fast changing workload and 30 seconds is no longer trivial anymore and what is even worse is on the primary subcluster side. Because the commit is much longer than five minutes let's assume, then at the point, you are committing to six nodes subscription all other nodes already waited for 30 minutes for GCLX or we know the global catalog lock, and the Vertica will crash the nodes, if any node cannot get the GCLX for 30 minutes. So the end result is that your whole database crashed. That's a serious problem and we know that and that's why we are already planning for the fix, for the 10.0, so that all the subscription will be batched up and all the nodes will commit at the same time concurrently. And by doing that, you can imagine the primary subcluster can finish commiting in five minutes instead of crashing and the secondary subcluster can be finished even in seconds. That summarizes the highlights for the improvements we have done as of 10.0, and I hope you already get excited about Emerging Eon Deployment Pattern that's shown here. A primary subcluster that handles data loading, ETL jobs and tuple mover jobs is the backbone of the database and you keep it running all the time. At the same time defining different secondary subcluster for different workloads and provision them when the workload requirement arrives and then de-provision them when the workload is done to save the operational cost. So can't wait to play with the subcluster. Here as are some Admin Tools command you can start using. And for more details, check out our Eon subcluster documentation for more details. And thanks everyone for listening and I'll head back to Dave to talk about the Eon on-prem. >> David: Thanks Yuanzhe. At the same time that Yuanzhe and the rest of the dev team were working on the improvements that Yuanzhe described in and other improvements. This guy, John Yovanovich, stood on stage and told us about his deployment at at&t where he was running Eon Mode on-prem. Now this was only six months after we had launched Eon Mode on AWS. So when he told us that he was putting it into production on-prem, we nearly fell out of our chairs. How is this possible? We took a look back at Eon and determined that the workload isolation and the improvement to the operations for restoring nodes and other things had sufficient value that John wanted to run it on-prem. And he was running it on the Pure Storage FlashBlade. Taking a second look at the FlashBlade we thought alright well, does it have the performance? Yes, it does. The FlashBlade is a collection of individual blades, each one of them with NVMe storage on it, which is not only performance but it's scalable and so, we then asked is it durable? The answer is yes. The data safety is implemented with the N+2 redundancy which means that up to two blades can fail and the data remains available. And so with this we realized DBAs can sleep well at night, knowing that their data is safe, after all Eon Mode outsources the durability to the communal storage data store. Does FlashBlade have the capacity for growth? Well, yes it does. You can start as low as 120 terabytes and grow as high as about eight petabytes. So it certainly covers the range for most enterprise usages. And operationally, it couldn't be easier to use. When you want to grow your database. You can simply pop new blades into the FlashBlade unit, and you can do that hot. If one goes bad, you can pull it out and replace it hot. So you don't have to take your data store down and therefore you don't have to take Vertica down. Knowing all of these things we got behind Pure Storage and partnered with them to implement the first version of Eon on-premise. That changed our roadmap a little bit. We were imagining it would start with Amazon and then go to Google and then to Azure and at some point to Alibaba cloud, but as you can see from the left column, we started with Amazon and went to Pure Storage. And then from Pure Storage, we went to Minio and we launched Eon Mode on Minio at the end of last year. Minio is a little bit different than Pure Storage. It's software only, so you can run it on pretty much any x86 servers and you can cluster them with storage to serve up an S3 bucket. It's a great solution for up to about 120 terabytes Beyond that, we're not sure about performance implications cause we haven't tested it but for your dev environments or small production environments, we think it's great. With Vertica 10, we're introducing Eon Mode on Google Cloud. This means not only running Eon Mode in the cloud, but also being able to launch it from the marketplace. We're also offering Eon Mode on HDFS with version 10. If you have a Hadoop environment, and you want to breathe new fresh life into it with the high performance of Vertica, you can do that starting with version 10. Looking forward we'll be moving Eon mode to Microsoft Azure. We expect to have something breathing in the fall and offering it to select customers for beta testing and then we expect to release it sometime in 2021 Following that, further on horizon is Alibaba cloud. Now, to be clear we will be putting, Vertica in Enterprise Mode on Alibaba cloud in 2020 but Eon Mode is going to trail behind whether it lands in 2021 or not, we're not quite sure at this point. Our goal is to deliver Eon Mode anywhere you want to run it, on-prem or in the cloud, or both because that is one of the great value propositions of Vertica is the hybrid capability, the ability to run in both your on prem environment and in the cloud. What's next, I've got three priority and roadmap slides. This is the first of the three. We're going to start with improvements to the core of Vertica. Starting with query crunching, which allows you to run long running queries faster by getting nodes to collaborate, you'll see that coming very soon. We'll be making improvements to large clusters and specifically large cluster mode. The management of large clusters over 60 nodes can be tedious. We intend to improve that. In part, by creating a third network channel to offload some of the communication that we're now loading onto our spread or agreement protocol. We'll be improving depot efficiency. We'll be pushing down more controls to the subcluster level, allowing you to control your resource pools at the subcluster level and we'll be pairing tuple moving with data loading. From an operational flexibility perspective, we want to make it very easy to shut down and revive primaries and secondaries on-prem and in the cloud. Right now, it's a little bit tedious, very doable. We want to make it as easy as a walk in the park. We also want to allow you to be able to revive into a different size subcluster and last but not least, in fact, probably the most important, the ability to change shard count. This has been a sticking point for a lot of people and it puts a lot of pressure on the early decision of how many shards should my database be? Whether it's in 2020 or 2021. We know it's important to you so it's important to us. Ease of use is also important to us and we're making big investments in the management console, to improve managing subclusters, as well as to help you manage your load balancer groups. We also intend to grow and extend Eon Mode to new environments. Now we'll take questions and answers