Sponsored by: Google Cloud | Scaling Data & AI with Efficiency: Databricks on Google Cloud Axion

Summary

Google Axion (C4A) is Google's custom Arm-based CPU that delivers up to 65% better performance than previous-generation x86 VMs, and Databricks' Photon engine is benchmarked to run significantly faster on C4A instances.
LSports, processing over 300,000 live matches per month, migrated from N2 to C4A instances on Databricks and demonstrated measurable anomaly detection job performance improvements alongside Unity Catalog governance adoption.
Internal Databricks data shows 54% of queries run at least 25% faster on C4A, legacy x86 migrations achieve 30–50% median performance gains, and serverless Databricks offers 63 regions with dynamic sizing on top of Axion.

Chapters

FAQs

What is Google Axion and how does it improve Databricks performance?

Google Axion (C4A) is Google's custom Arm-based CPU that delivers up to 65% better performance compared to previous-generation x86 VMs according to Google Cloud. When paired with the Databricks Photon engine, the combination delivers breakthrough price-performance efficiency for analytics and machine learning workloads.

What results did LSports see after migrating to Axion on Databricks?

LSports, which processes over 300,000 live matches per month on Databricks and Google Cloud, migrated its workloads from N2 to C4A instances and benchmarked improvements in anomaly detection job performance. The migration also included adopting Unity Catalog governance alongside the compute upgrade.

How much faster are Databricks workloads on C4A compared to x86?

Internal Databricks data shared in this video shows that 54% of queries run at least 25% faster on C4A, and migrations from legacy x86 instances achieve 30–50% median performance gains. These numbers reflect actual production workload measurements across Databricks' customer base.

What are the trade-offs between serverless and classic Databricks on Axion?

Serverless Databricks on Axion offers dynamic sizing, the latest Photon version, and availability across 63 regions, along with faster cold starts due to its auto-scaling architecture. Classic clusters offer more configuration control, but this video highlights that serverless trade-offs are manageable for most workloads given the performance and operational benefits.

Full transcript

[00:08] Good morning, everyone. Nice to see you all. I'm Varun Shah. I am a product manager at Google Cloud. I'm responsible for the C-series VM based on Google Axion, which is our
[00:22] custom arm-based CPU. Together, I have my partners El Sports. So, we have Zebic, who will be talking about how
[00:34] they're using Databricks on Google Cloud to accelerate the performance while being efficient. And I have 10 from Databricks side who will be talking about how
[00:48] Databricks is using the Google innovation to make Databricks faster and efficient. So, without further ado, let's get started. Everyone in this room, I bet, can talk
[01:02] for hours how AI is evolving their industry, disrupting their industry, and evolving your day-to-day workflow. Now, at Google Cloud, we are constantly thinking about how we can optimize our
[01:17] infrastructure to to scale the core performance without really breaking the bank. We want our infrastructure to be AI-ready for what's coming next. And we also want to make
[01:29] sure that we are aligning consumption to the workload needs, optimizing resource access through declarative and flexible automations to align the consumption to
[01:41] different workloads. Now, the infrastructure that we are designing, of course, we want it to be AI-ready, but also we are also making sure that the cloud first workloads,
[01:53] which are the LEGO for delivering this AI workloads, is not forgotten. We are also ensuring that enterprise first workloads, which are the bedrock of running day-to-day business for our
[02:05] customers, uh are taken care of and the key tenants of security, reliability, and uh some of the other tenants are well served.
[02:19] Now, the heart of our strategy is the Google compute platform, which is workload optimized, TCO efficient, and enterprise ready. It is powered by a technology called
[02:31] Titanium, which is nothing else but a system of offloads that helps remove some of the common bottlenecks of compute, networking, and storage. It is used by Google for decades to
[02:44] deliver planet-scale services like Gmail, YouTube, Search, and we want our customers to use uh the platform powered by Titanium to serve like Google.
[02:57] Now, what customers use is GCE, which is our uh VM instances. We offer flexibility of general-purpose and specialized.
[03:09] Now, GKE provides the orchestration layer, and it has something called the custom compute classes uh that enables our customers to use the compute
[03:21] based on the workloads they need and the consumption model they need. Paired with all of this is Hyperdisk, which is our newest generation of block storage, which enables our customer to
[03:34] provision the throughput and IOPS as per their needs. So, the net-net is over the last few years, we have been working to expand our portfolio of VMs
[03:47] to offer workload optimized compute. Now, on the left, what you're seeing is the general purpose compute and on the right, we have the specialized compute. Now, in the general purpose side, we
[04:00] have two offerings, which is our C series offering, which is the performance workhorse for your business. It provides the It provides the different features such
[04:13] as Hyperdisk as well as the local SSD, tier one networking, and advanced maintenance. Think of C as the performance workhorse for any of your workloads. And we have
[04:27] the N series, which provides you a balance of throughput and performance. Now, on the specialized compute, we have various offering. We have the A series, which is
[04:39] enabling your HPC style workloads. Then we have the Z series, which offers the storage optimized, giving you a higher local SSD to VCPU ratios. We have the memory optimized, which enables you
[04:54] to get much more memory per VCPU. And we recently announced the network and block storage optimized VMs at Next in April that offers much higher networking
[05:07] throughput as well as the storage performance. And on the right most, we have the accelerator optimized, which is nothing else but our GPU lineup as well as our lineup of TPUs, including the latest 8i
[05:22] for your low latency inference workloads as well as the 8T for your training workloads. So, net-net, we are offering our customers a choice to pick the right
[05:34] tool for the right job and and providing a workload optimized portfolio. Now, what you're what you will see is our customers use majorly
[05:47] the general purpose compute and that is where we are innovating with our Google Axion which is the arm-based custom CPU. Now Google Google Axion is nothing else
[06:00] but based on the arm ISA. It provides up to 65% better performance than the comparable x86 current generation VMs. Now we are able to do all of this at a 60% better
[06:14] efficiency. And today in the AI era power is of the utmost importance and so this is something that we are all bringing at the right time to help with that
[06:26] situation. The great thing about Google Axion is it is standards based. That means if you're running arm elsewhere, you can lift and shift and it has a broad ecosystem support
[06:39] that has been evolved for the last decade and you can really go and adopt arm with little or no code change. Now the number speak for itself.
[06:52] We are seeing up to 30% better Java and MySQL performance and we are able to get all of this at a 60% better energy efficiency. Now with respect to the VM offerings, it offers we offer up to 72
[07:06] vCPUs, 576 gig of RAM. That is you can get up to 8 gig of RAM per vCPU. It comes with the Titanium local SSD. Now again to remind Titanium is the system I
[07:19] spoke about that helps you remove some of the common bottlenecks of compute network and storage and gives you more compute, low latency local SSDs and hyperdisk. It comes with 100 gig of tier one
[07:31] networking and with standard it is up to 50 gig. It also comes with hyperdisk all flavors balance throughput and extreme. And again it is based on the latest generation of the titanium
[07:43] offloads. Now, this VM has been out for roughly 18 months and we have heard great things. Customers are thrilled to receive this VM. They're adopting it at a large scale and their
[07:58] usage we are seeing is various based on customer profile. Customers are using it for web applications. They're using it to run the back-end database, analytics. Uh
[08:10] they're also using for media encoding, gaming servers, EDA, and so on. One key thing is they are able to validate the value proposition, which is price per of this
[08:22] product, and so we we are thrilled to announce the N4A earlier this year, which which brings the price performance even further uh for our customers.
[08:35] Now, not to mention not to forget that Epyc is AI ready. That means you can run the agentic sandboxes, small model inferencing, the AI pipeline processing, vector search, and so on.
[08:50] Now, we have three uh family three VM uh families available in this portfolio of Epyc, uh C4A, N4A, and we also recently announced the GA of C4A metal, which
[09:04] lets you have a direct hardware access to the C4As. I also want to highlight that uh we we recently announced the preview of Z4D,
[09:16] which is our high-performance storage and networking VM. And again, it comes with 84 TB of titanium local SSD, 400 gig of networking. It It is based on the uh
[09:29] latest generation of AMD Epyc Turin CPUs, comes with 384 vCPUs, so giving you the compute to scale out with 3 TB of RAM giving you enormous memory to run in-memory analytics and so on. So, if
[09:43] you're interested, please feel free to scan this QR code and sign up for the preview. Um It's a It's a great product. Now, with that, I want to invite over
[09:56] our colleague from L Sports to talk about how they're using Databricks on GCP.
[10:10] >> Okay, can you hear me? Hello all. Uh my name is Zevik. Zevik Gershner. I'm a group manager of Product House and L Sport. L Sport is one of the biggest sports data real-time sports data
[10:23] providers. We are providing I will show you in a minute. We are providing all type of sports data in-play or post-match, pre-match, etc. And
[10:37] >>
[10:49] for sports books. It's It can be the the the odds, the fixtures, etc. And Our challenges is
[11:01] is deliver live data with low latency during during a big match like the World Cup that we all love and enjoy to see now. Uh we need to
[11:13] to detect and catch uh anomalies on the data on the real-time data as fast as we can and keep the settlement etc. mapping accurate.
[11:28] And of course, we need to control the compute cost uh because of you will see in a minute that we have a lot a lot of data and it's scaling and scaling and we cannot
[11:41] lose no no the latency or the performance and no the cost of course and we have a list of layers that we must
[11:53] be on sport so in a glance we have around 300 K live matches fixtures covered per month for all type of sport 70 70 types
[12:08] from from football or soccer basketball to kabaddi and badminton etc. and for each match we are providing odds and
[12:21] markets for the bookies it's about 11 million markets but if we will go drill down to the bets themselves and the odds we are talking about
[12:33] billions because each market has a lot minimum of two or three bets and for the anomalies that we need to to catch up fast as we could for now we
[12:45] are our SLA is 5 minutes for anomaly the latency itself of the feed in in the in the in the in play itself if we want to provide the statistics corners goals etc. etc. we are talking about
[12:58] less than a second around 300 500 milliseconds per per update from the field
[13:11] so till 25 we still we we we didn't worked on GCP we worked on other cloud cloud provider and we saw that after our benchmark etc.
[13:23] that if we We move to GCP to Google Cloud. We will benefit from the infrastructure and the cost. So, we moved on and consolidated all to GCP. It
[13:35] took us 1 year end-to-end, including databases uh hundreds of clusters and nodes, etc. Then, we moved from our previous uh
[13:47] uh data platform we that we used to uh Databricks, leveraging the um Unity Catalog, the governance, uh etc.
[13:59] And um in the last couple of uh uh of months and weeks, we are shifting from using our old type of uh machines.
[14:12] Most of our machines on Databricks was N2. And after benchmarking that we'll show you, we moved to C4A and we are migrating all to there. So,
[14:26] from my level, how our system is work, we are working with Bringg, that it uh it's a company uh inside company that's uh pro- that bring in any da- data that that
[14:40] they can from the free free worldwide web. Uh with a low latency, and then we are going we are pipelining with uh Kafka.
[14:52] Moving to ingesting all to the uh to the lakehouse with Spark. And for the real-time data and the real-time analytics,
[15:04] the because just only only today I saw I heard that Databricks has a has a solution for it, but they didn't add till ye- till till till now, so we are using uh ClickHouse. And all of this
[15:18] wonderful uh pipeline ends up with the Arena with all Huawei uh products.
[15:33] Okay, for for So, you you see I talked about it. We are using the Unity catalog. We have also Delta tables and iceberg tables. And we are using a Photon for a lot of our
[15:45] heavy lifting jobs for better performance and latency. So, I will show you now a two different type of jobs that we are running on L sports
[16:00] that we migrate to C4A. The first one is a weight sale pipeline. What it does, it take any history a lot a lot of history of live statistics
[16:14] they while was live previously and with the algorithm and machine learning models, we are ranking each statistic
[16:26] per provider, per scout based on the history. So, when we have a live match, we can choose and decide what is the
[16:40] real decisions that we need to make that happened in the in the field based on the provider quality and ranking. So, it's uh it's offline. Yeah, it's
[16:52] offline job that used in the real time in the real time pipeline. Um it's really it's kind of a heavy because we have a lot of months as you can see,
[17:07] a lot of reads and uh upsert etc. The other The The job that I will show you we benchmarked is a anomaly detection
[17:21] inference. This is a real-time job that uh needs to to detect faster
[17:33] an anomaly in the in the statistics in the data like let's say if a provider is now sending wrong odds or outlay odds etc. Or the statistic is not make sense etc. So etc. So for now
[17:47] we are talking about uh seven type of anomalies that we have a 5-minute SLA um and 14 task in parallel. This is a lot of
[17:59] group by and aggregations and windows and joining etc. So we have two different type of jobs. One it's offline. One it's online.
[18:12] Um and you can see here that when we benchmark both of these type of jobs against N2 or C4 to C4A
[18:26] we you can see that X on the C4A was faster in both of them. Per the job itself and even per a task in a job which is a broadcasting group
[18:38] by window function you can see that in the in those task it was 2 2.5 faster than the N2 the older generation.
[18:51] For the migration itself for for this type of jobs we didn't do anything except just let's run it on the C4A. Yeah any we didn't Here we didn't did any changes in
[19:05] the code or in the pipeline etc. So we benchmark the serverless and Photon in the classic to see
[19:17] how much it's cost-effective and the performance-wise, etc. So, we see that without a Photon we are nearly in the in the we are
[19:29] missing the SLA. It's the 5 minutes of the anomaly. So, we must use the Photon here and accelerate the job and eventually we are bringing our customer what they need.
[19:42] And in the classic run without a serverless so it's it's run faster than the the serverless. Well, my guess it's because of the cold start and etc. But
[19:57] even the the serverless with a Photon it was in the SLA. Almost a breach it, but it didn't breach it. So, you you can see here that Photon
[20:09] is a must for for the heavy Sorry for that. For the the heavy lifting of windows and the aggregation group buys, etc. And the classic is better if you can't tolerate
[20:23] the penalty of the latency of the cold start. So
[20:35] for every changing on node types, we are suggesting warmly warmly to benchmark first the benchmark the the performance, the cost, and yeah, the cost is
[20:48] is 20 10 20% cheaper than the regular chips. And the performance, as I said before, as I showed you before, it's 20 to 24%
[21:01] higher in the latency, but for a real benchmark for for for before you are moving, I really suggest you to buy the identity card the the the output
[21:13] to verify that all is good and and and in place and just then to move to the action C4. A. And that's it.
[21:26] Thank you very much. >>
[21:42] >> Right, so hi everyone. My name is Tan John. I'm engineering manager working in the compute platform organization. You know, my team is mainly focusing on orchestrating computing resources for all the data bricks workloads. This
[21:56] includes like you know, data analytics use cases, SQL warehouse and you know, bunch of like the lake flow use cases we talked about in the keynote. And also like even like the new reader
[22:08] engine we talked about in the keynote is actually going to work basically you know, on the orchestration platform, you know, basically my team is running. So we work very closely with like all the data bricks like product teams, you
[22:20] know, on one hand I'm empowering like customers, you know, like like you know, L sports and supporting them to grow the business. But on the other hand, we are actually working very closely with Varun, you know, on teams and you know, other cloud providers
[22:33] to basically optimizing the underlying compute resources. Um and you know, at a very high level right, so you know, at the data bricks products right, we're focusing on data analytics as you all know.
[22:45] And also like you know, our specific focus one of these lines actually SQL warehouse and we'll talk of more about like our data, but it's actually very growing fast like business. And and then the third one is basically
[22:57] you know, Unity catalog. As I've been talking about that actually a lot of customers actually like like facing the centralized like a data governance here. Uh and on top, right? We talked a lot about like, you know, 1 Genie, right? So, in the keynote today and a lot of AI
[23:09] BI applications actually built on top. And all of these actually require very cost-effective and fast uh hardware. Um Uh and you know, as we actually exploring different uh compute hardwares, and in
[23:22] fact, actually that's a dominant factor of our compute platform, we realized actually C4A works uh really really well. And you know, we observe there's like 29% uh medium performance gain
[23:34] compared to like, you know, our old second-generation instances, uh which I've been using like serverless and the classic. Uh and the main reason is that we noticed actually the processors is actually way more cost-effective and
[23:46] both in terms of like VCPU cores, but also actually associated like titanium uh memory, like uh and also actually uh even like the the remote disk we use like sort of hyper disk balance mode.
[23:58] So, you know, both actually remove like all the you know, a lot of these IO uh bottlenecks. And that's what we see actually the main gain come from. And we have like more data to talk about later. Uh and we also see actually it works really well with like AI-based uh use
[24:11] cases, you know, uh and low-precision floating-point and a bunch of other like really nice features there. Uh and also like uh security is actually uh quite nice, you know, with like security and micro uh controllers and encryption at
[24:23] rest. Uh a lot of these sometimes are asked by like, you know, our customers, especially for those are very like uh you know, uh secure uh customers. Um Um and then recently, you know, given this like very good like performance of
[24:36] C4A, we started migrating it to basically all of our serverless uh DBC workload uh on this new chipset. Uh and this graph literally shows like, you know, the performance uh improvement
[24:48] here. Uh and you know, as you can see, the the X axis actually is a little bit hidden, but basically what it shows is like, you know, the first part is about like what is the number of queries we see improvement. We see about like 89%
[25:02] of the queries after migrating C4A, we see improvements. Uh and the second column shows like, you know, there's a 54% of queries uh that are at least 25% faster.
[25:14] Uh and the third column actually shows actually about like uh 20% of queries that are two times faster. And we see only like uh 6.7% of queries that remain performance neutral. So, this actually
[25:26] demonstrate this new chipset actually have significantly benefited like most of our uh DBC core workloads. Uh and in fact, we are migrating like a more like Lake Flow like you know, other products from Lake Flow uh to C4A. Uh it's actively
[25:40] ongoing. Um And we also have like a more for like, you know, drill down more on like why there's a performance improvements. Uh we break down the data based on like the size of the queries using the same data
[25:52] side. Uh what we observe is that the bigger the queries actually we see even higher performance improvements. Um and for example, if you see actually if
[26:04] the queries actually last like more than 5 seconds or at least or even longer, we see actually performance improvement is at least uh 30% and sometimes for even bigger queries can go to 40%. Uh and the main reason there is, you know, it's
[26:18] actually a lot of this is because of the storage improvement in Titanium uh and the balanced hyperdisk mode. Um
[26:30] And we also actually look at like basically what type of workloads can, you know, be benefited uh you know, differently from C4A. So, these are actually benchmarks actually showing uh the how does the C4A compare to our previous like third generation uh
[26:42] benchmarks. Uh and the X axis it shows like uh what is the percentage of like performance improvement. So, you know, bigger the better. And Y axis shows like different like benchmarks.
[26:54] Uh and what we can see is actually the first two, TPCDS and TPCH, are like typical like sort of like data science like analytics workload. Uh we see about like 30 to like up to 50% actually a performance improvements. Uh and the
[27:07] middle three are more about like basically like what is the performance in terms of like shuffle intensive like data intensive workload. And we see actually really good performance, at least like 30% increase. Uh and even for like, you know, heavy like storage
[27:20] workloads on the two bottom side, we see at least about like 25% improvements. Uh so this data set actually shows C4A actually benefits different types of uh data analytics workload uh really well.
[27:33] Uh so you know, with that information in mind, so you know, we we have the following suggestions for our customers to use C4A uh in classic compute. Uh the first one is uh please considering
[27:45] compounding the advantage. Uh and by compounding what I mean is, you know, we can leverage C4A, right? It's like faster processing cores, better IO, fast networking. But also trying to use the latest Databricks
[27:57] runtime like 8.x because it has a lot of like, you know, improved uh features such as auto scaling, uh Photon, and like, you know, and latest Java libraries, you know, all of these will actually, you know, collectively
[28:10] improve performance. Um and the second one is, you know, we we advocate to customers using classic compute to upgrade to the first fourth generation like C4A chip set. Uh you know, we notice a lot of
[28:22] customers still use legacy x86 Gen2 instances. Uh and you know, in the previous slide we showed there's at least like 30% media performance gain and sometimes they go up to 50%.
[28:34] Um and one caveat there is, you know, uh please make sure we you get the sufficient C4A quota before migration. Uh and also C4A, you know, because it's actually a very good uh chip set, so
[28:47] sometimes, you know, the capacity in certain regions not sufficient and there, you know, it's good to implement like four by strategy and the data breaks we have a four by strategy called flex node types. If you enable
[28:59] that in classic compute, it can automatically find like, you know, a four by instance type that can also provide good performance. Uh and lastly, you know, we think it's
[29:11] actually very important to leverage hyper disk well, especially tune the throughput IOs accordingly based on the VCPU core. That's what I see the most performance improvement. Sometimes, you know, can go to 60%.
[29:24] Uh lastly, if you if you don't like a lot of this hand tuning, we also advocate our customers to go with just serverless offerings. The benefit there is it's actually
[29:37] automatically manage all the underlying compute capacity. So you don't need to worry about like quota and you know, the transient capacity issues always actually handled by the serverless platform. Um and also like we have internal
[29:50] optimizations on safeguards for making sure, you know, even there's like a transient capacity issues, quota issues, the system can bridge the gaps. And we also actually are actively working on cross region serving to support like better failover.
[30:03] Um and also like, you know, with the serverless, it can automatically pick like, you know, even better auto scaling algorithms and it also dynamically sizing the compute nodes based on your
[30:15] workload characteristics and IO demand. So it can further optimize actually the performance. And also like, you know, it always use the latest and the best runtime configurations such as the latest Photon version and like, you know, data bricks
[30:28] runtime upgrade. Uh and also like for serverless, we already have like 63 regions coverage and offering like two different modes like performance optimized one and the standard mode. So you can actually balance different cost and latency
[30:42] requirements. Uh and we are actually expanding to like more regions, you know, over this year. Um and also we are actually have a demo on tomorrow like 3:30 which talk about even if you have a classic workload, we have
[30:55] like to build the toolings and the agents that can automatically migrating classic workload to serverless, you know, just to streamline sort of everything there. Yeah. Okay, that's the end of my talk. Yeah, thank you everyone.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.