SIGMOD Systems Award Talk on the Aurora Database System

Video in TIB AV-Portal: SIGMOD Systems Award Talk on the Aurora Database System

Formal Metadata

Title
SIGMOD Systems Award Talk on the Aurora Database System
Title of Series
Author
License
CC Attribution 3.0 Germany:
You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor.
Identifiers
Publisher
Release Date
2019
Language
English

Content Metadata

Subject Area
Loading...
Service (economics) Group action Service (economics) State of matter Magneto-optical drive Relational database Database Database Mereology Element (mathematics) Power (physics) System programming Self-organization Physical system Identity management Point cloud
Source code Theory of relativity Open source Relational database Multiplication sign Relational database Database Database Replication (computing) Software maintenance Cartesian coordinate system Element (mathematics) Workload Type theory Different (Kate Ryan album) Core dump Point cloud Energy level Integrated development environment Right angle Physical system Point cloud
Scaling (geometry) View (database) Database Shape (magazine) Replication (computing) Mereology Database transaction Mechanism design Different (Kate Ryan album) Electronic visual display Software framework Multiplication Area Enterprise architecture Channel capacity Relational database Software developer Data storage device Cloud computing Maxima and minima Sequence Wave Different (Kate Ryan album) Pattern language Point (geometry) Dataflow Server (computing) Service (economics) Channel capacity Virtual machine Maxima and minima Data storage device Prime ideal Architecture Workload Internet forum Band matrix Computer hardware MiniDisc User interface Addition Scaling (geometry) Consistency Forcing (mathematics) Model theory Basis <Mathematik> Database Stack (abstract data type) Cartesian coordinate system Integrated development environment Computer hardware Point cloud
Scaling (geometry) Source code Database Instance (computer science) Client (computing) Mereology Replication (computing) Stack (abstract data type) Database transaction Perspective (visual) Mathematics Semiconductor memory Core dump Social class Identity management Point cloud Enterprise architecture Channel capacity Data storage device Database transaction Bit Instance (computer science) Replication (computing) Disk read-and-write head Connected space Type theory Process (computing) Workload Universal product code MiniDisc Pattern language Data storage device Power (physics) Workload Cache (computing) Reduction of order Medizinische Informatik Backup Mathematical optimization Scale (map) Addition Client (computing) Database Line (geometry) Cartesian coordinate system Single-precision floating-point format Software Integrated development environment Point cloud Abfrageverarbeitung Identity management Local ring Spectrum (functional analysis) Library (computing)
Context awareness Equaliser (mathematics) Source code Workstation <Musikinstrument> Execution unit Database Client (computing) Parallel port Disk read-and-write head Benchmark Different (Kate Ryan album) Core dump Multiplication Identity management Channel capacity Decision theory Web page Median Database transaction Instance (computer science) Demoscene Process (computing) Gaussian elimination Workload Buffer solution Order (biology) Router (computing) Web page Point (geometry) Slide rule Sequel Dependent and independent variables Computer-generated imagery Control flow Data storage device Event horizon Rule of inference Scalability Product (business) Number Crash (computing) Computer hardware Selectivity (electronic) Backup Scale (map) Scaling (geometry) Physical law Content (media) Computer network Basis <Mathematik> Volume (thermodynamics) Cartesian coordinate system Software Personal digital assistant Predicate (grammar) Query language Abfrageverarbeitung Code Multiplication sign Insertion loss Instance (computer science) Analytic set Mereology Component-based software engineering Befehlsprozessor Decision support system Query language Process (computing) Logic gate Point cloud Curve Data recovery Real number Data storage device Benchmark Disk read-and-write head Connected space Type theory Right angle Volume Resultant Reduction of order Data buffer Backup Service (economics) Channel capacity Data recovery Virtual machine Lemma (mathematics) Workload Mathematical optimization Condition number Shift operator Cellular automaton Database Predicate (grammar) Scalability Human migration Point cloud Pressure
my name is the state funny a distinguishing Jr. in a
WSN down 1 of the members of the team that built power or and I'm grateful for the award of today as you can see
this was a bigger group that for this is just actually the names of a few of the people working on the service so it it was a big investment for KWS and the database organizations that I am part of it the U.S. OK so the state is this that I'm giving today very briefly is is based on the tool of the person that we published in the last few years said about Aurora there were published era for a sick mother so
that that encourage you to go back and read them if you haven't yet to be closer to the core over what amount about to say in much more
detail than I can cover with the time that we have right now the rigor so when we start to building our or our we had the a fundamental goal of which was Boole redesign relation of that the Blazers for the cloud that that was the level goal and all entered the BIO so we did the same thing go without a database systems so like Dynamo dB for ownership of the Cabezas and the redshift there for analytical workloads for we wanted to reinvent the relational databases set and at the same time we wanted to on the net backwards compatibility with the most popular open-source engine so that exist that and the reason why we wanted to do that is because a especially my see column POSCO a very wide the adoptions of both in academia and the industry and we realize that there converting their applications to run on different types of engines is a big big article for customers because you basically have to read design your replication so 1 of the fundamental tenets of for our oral was to maintain fullback cost compatibility with the engines tool enable what we call the lift they shifted the ability to take a database and application ported to the cloud ported to Aurora the benefit from the advantages so that gives us which are mostly around the performance and durability availability and the scaling and so we started with the Aurora Mexico compatible demands 2014 the followed by post-processing in 2017 alright right and this is just the
feature of the the launch party for our in 2014 the forum a sequence of only Washington you see OK
so we talked about how Aurora the goal of reducing by designing their relational databases for for CloudBees architecture so but how Aurora started the development that a few years after the launch of the WSM by then I think we had already I did the grasp of what the the new wave of cloud-based work also look like at that point that so a AWS was all about the basically providing have very flexible hardware free said they were based on commodity hardware at to basically the able to give a very low cost for customers on top of that we also noticed that by running the server workloads as part of our our cloud infrastructure the fault of war workloads are very bursty do the use of them burst of the traffic of full of sometimes by you know much more traffic patterns so where the traffic is it is not as high as the peak so traditionally now prime user environments that is necessary to provision for peak traffic of because of the capacity the hard work a price at the it's is not flexible but this basically that forces you fold and higher cost of because we need the provision for the maximum throughput it your replication is to sustain on the other hand that the cloud that which is basically a very large and flexible full of hard work allows to tool design the fleets that provide the on-demand capacity when you need it the and take this capacity away when you don't need any longer though a mechanism that we call automatic scaling which is basically about ending it and removing capacity to I a pool of machines so that that supports some application that there in the reaction or proactively that 2 ships in the war flow in the workload trumpet shape but unfortunately this is harder to do with that the basis of because there is so much easier to wear than remove capacity when you have a pure compute fleets that don't have to worry about you know durability and consistency that creates a much bigger challenges when it comes to keeping the fundamental promises of the the basis about consistency and durability and available to another trait of cloud-based architecture is multi tendency the is the fact that that that weaker located the framework also belonging to different customer service to you on the same physical hard work I was doing that in a very secure fashion the reason being that there's a lot of our work closer bursty but the burst are not always the same difference Croats from different customers in different patterns so the ability to collocate them together on the same model allows you to leverage the were bad and the Steelers reducing cost because again we don't provision for peak usage they good the so let's look at some of the area and point of view and again the the the papers that they mentioned going much more detail into this this look at that additional database that that additional database that usually as a locally attached storage and because of the additionally your display you always been the bottleneck of for a database applications and so it was necessary would be necessary to have a locally attached storage to maximize the you bend with the minimizer of at least 6 and this is basically a pattern that has lasted you know for 40-plus year however in the club is environment that it brings a lack of flexibility because we know will collect that storage again is something that is very esthetically provision is not very elastic it is not very flexible so the way we have
really rethought the database in the cloud architecture that is basically by for and you were trained that which is basically to decouple compute the and the storage of and the reason is the four-winged that we want to optimize a and scalar the tool but power portions of the database stack of what we call on the up perspective which is more about you know connection management transaction processing and so on and the bottom part that which is more about the storage of durability and the availability of the data so by decoupling the computer and the storage and basically connecting the donor network and were able to optimize them independently to war after the core tenets that each of them each of the 2 parts of this spector basically needs to satisfy the writer said that the storage then becomes so optimized for durability and takes care of replication for high durability backup-and-restore where we can also optimizer the computer Lee you're the 1 on top of to be flexible as workloads changes so for example a patterns that we see customers' shut down the database instances where there I because they know that the data face at durable in the storage and they just in the up spin then down when no longer needed and this is about in a cost reduction it is possible to the computing as fast as a tool the to basically skilled them up or down to larger or smaller instances with more memory or less memory more secure less if you again to adapt that work what changes and optimizer for and then the social possible chorus of course so really because of for example rescaling and again it's possible to add or remove read right because for from the class that the basically provided additional re capacity the bottom line is that we have now converted from locally attached storage of tool our remote storage though the network and is a shifted the the bottleneck to the network itself as opposed to the local disk aerial assault of the engineering work it has been done in our was about optimizing the the network access patents and make them out best stuff for the type of environment indeed AWS fabric very good
let's take a quick look at the 2 layers of the stack so like we said that the top half of the spectrum which is the computer layer but user B is there on a single use this is so in Aurora Bay instances that can connect to the customer application that they're actually single-tenant that they're up they are applied that to individual customers to intervene we knew there were close and the cooler the core tenet the forward the the top of part of the spec is to basically maintain full compatibility with the engine so the I see for Pospisil down to the wire problem also for example is possible to use exactly the same as the client libraries so that you would use with my Sihpol or post of source and then attend also so that additional duties of the upper half of a bit of respect like an money management the query processing transaction management of manager but for manager and so on improving the the undo the in memory on the other hand the the aurora storage
at the sports component is the 1 that is most the novel in different if you will and basically the or stores component that is defeat of machines so that we manage of behind the scene on behalf of the customer that is multi-tenant that the user equal different workloads and different that the basis from different customers this freak that user scalable that uses possible to add or remove capacity is needed that i it's replicated that to guarantee availability and durability and the cell feeling which means that the as instances of your work break care or become inaccessible we can add them more tool replace them and this is important because to drive down costs I need a BIOS would be is aware hardware freedom commodity hardware which means that the instance loss of a something that we need to deal with than that tolerate in a transparent fashion so with that in mind what happens then is that that the aurora storage is responsible to replicate the data and the application user Sixways ways across the different the percentage that we call availability source and there is only implicate the 6 ways is for basically make sure that we can always tolerate that would at the center outages said in case of a situation like you know dramatic power outages or weather events and so on and then the backups about lemmas honestly OK
so how did we eliminate the network bottleneck they keep the core concept again that were discussed the to answer in the paper is still only pusher read locks so to the storage that another war so I like you know the the core with the course where you have a right had log and then you also then they're in the back on a synchronously solutions that 2 pages from the buffer pool of you know the head node that the top layer only pushes redo log to their storage and then the storage component that continuously performs recovery that is the code that you normally would run run after a pressure and performance continues it performs back but this allows us to perform a real recovery in this baby fashion across the direct because of the storage volume and this came about that so what this means is that the head node that does not need to push the dirty pages which means that this greatly decreases the amount of network that your that we need to push down the to the to the storage of and this goes back again to optimizing network your because network your becomes the new ball on that the other interesting thing is that this is would perform a redraw or a recovery all the time and essentially fresh recovery becomes almost instantaneous which means that you can actually becomes much easier to tolerate that the Cabeza restarts of because in on that wait for a very long the crash recovery process that normally takes place so with right intensive workloads and this allows us to the flexibility also can be another component because it becomes easier to replace the ahead in order for example for scaling events for fear of conditions for patching that because again because very quickly to restart the the database engine on a different machine for the more the fact that we removed these your bottlenecks also allowed the story mover contention in a bunch of other places in the code that but the normaI you wouldn't you best thing because they were all contention boys there were masked by the that additional although makes that world that things to be so different parts they good the soul of the and will compete quickly give a couple of a performance so you station so this from a young workload that this was before and after the Aurora migration so you can see the difference in their elite and said that they were seeing and this is something that would publish it in the 1st of the 2 papers this as likely in your numbers since then there there be used on a fairly restore cis venture workload though so you can see that we have a only in right only were product and you can also see our it's running on different type of fire were bigger to the left as monitor the writer so I would call out a couple of things not just the the rule I know which is the aurora numbers higher but also the almost nearly with a linear scaling the compared to what my sequel does where you therefore selects a kind of have this curve at the factor and for right so it basically doesn't scale fervor beyond a certain point and these are obtained my support by 7 unit of U.S. but then that we talked about the fact that the aurora storage component essentially can out their rights things were not taken a fleet that recently also launched new feature which is polyp create which basically gospel read so what we originally did for writes a so the B is that since the storage as both the law and the and the Database pages because he continues around free tools we were also able to push the largest concept down to the fleet that especially for we use that they're very selective predicates or aggregations of because were basically able to leverage at the very very vast pool of machines so that the Aurora sports component is comprised of and for large analytical we use of viewable tool performed a large-scale extensive very distributed fashion and then you have you know that that gates the results on top of so this is basically a senior optimization that would publish the in the last year or so for for the intensive workloads and this is basically some slides that we have on our well known as a decision support benchmark published my Netflix said that what you see is basically this the speed up that they observed that for the individual creates with suffered one and then 20 x and greater than 10 x speed of for each of the 22 queries in the bank's market and is a large-scale undeletable queries that they were running their comparisons for against all right or or or or my sequel last slide that these I recently new features that we launched the is called were a serverless and is that most cloud 1st of feature that we have a lot for so our 7 this is about that taking away even the concept of a freaks the head node that computing instance so the way our service work said is that that we have a pool of her request routers the top proxying declined connections to the database and then the head notes are utterly added the removed them on the fly as the work product shifts because we have notice that there are lot of war cross with the database actually goes idle and something gazes median Eitel during the week and the ensuing over most of the fact that there is even a head the instance of the takes care of the transaction management like connection managers is completely extracted and so we have a war all of the instances of instances can go in and out of the War phone and that this new request component that that we developed is basically takes care of managing client connection and holding them up as instances coming from the cluster and this was my last slide this so I
want to thank you the committee for the award a very prestigious and we're very very grateful thank you
very much thank
Loading...
Feedback

Timings

  391 ms - page object

Version

AV-Portal 3.21.3 (19e43a18c8aa08bcbdf3e35b975c18acb737c630)
hidden