Uses and abuses of cloud data warehouses(materialize.com) |
Uses and abuses of cloud data warehouses(materialize.com) |
At the time we were mostly a batch data shop, based on Apache Airflow + K8S + BigQuery + GCS in Google Cloud Platform, with BigQuery + GCS as the central datalake techs for analytics and processing. We still had RT capabilities due to having also some Flink processes running in the K8S cluster, and also having time-critical (time, not latency) processes running in microbatches of minutes for NRT. It was pretty cheap and sufficiently reliable, with both Airflow and Flink having self-healing capabilities at least at the node/process level (and even cluster/region level should we need it and be willing to increase the costs), while also allowing for some changes down the road like moving out of BQ if the costs scaled up too much.
What they wanted us to implement what according to them was the industry "best practices" circa 2021: a Kafka-based Datalake (KSQL and co.), at least other 4 engines (Trino, Pinot, Postgres and Flink) and an external object storage with most of the stuff running inside Docker containers orchestrated by Ansible in N compute instances manually controlled from a bastion instance. For some reason, they insisted on having a real time datalake based on Kafka. It was an insane mix of cargo cult, FOMO, high operational complexity and low reliability in one package.
I resisted the idea until the last second I was in that place. I reunited with some of my team members for drinks months later after my departure and they told me the new CDO was already convinced that said "RT-based" datalake was the way to go forward. I still shudder every time I remember the architectural diagram and I hope they didn't finally follow that terrible advice.
tl;dr: I will never understand the cargo cult around real time data and analytics but it is a thing that appeals to both decision makers and "data workers". Most businesses and operations (especially those whose main focus is not IT by itself) won't act or decide in hours, but rather in days. Build around your main use case and then make exceptions, not the other way around.
Is the desire for "RT-based datalake" itself misplaced? Or just that the implementation isn't up to the job? Nobody _wants_ slow data, and reports that are usually fine with T+1 delay can become time critical (for example, a "what's selling?" report on black friday).
But as an engineering manager, at least I must ask and answer the following question: even when nobody wants _slow_ data, how fast is fast enough? I don't see decision makers choosing and thinking better with a 10 min latency vs 20 min latency, as they are not looking at the reports all the time, even for big events like Black Friday (they have meetings and stuff you know, even their supporting analyst teams do).
For more time-critical matters (i.e. real time BI or real time automatic microdecision making for fraud detection), as I said, we did have the capability to run both more frequent microbatches or do RT processing using Flink connected directly to our app backend messaging system (ironically Confluent, Kafka as a Service). But that is very different to using a complex real time log as Kafka running on "pet" servers as the cornerstone of your data platform and then propagating said data to different engines/datastores (at least 4 as I said) for downstream processing. That's a lot of moving parts running in a low reliability environment.
Overengineering is a thing, and I think it was my responsability at the time to limit the level of complexity considering the reality of the business and the resources we had in the team, even if that meant 20 minutes of latency for a business report. That's my point and why I say I think it was a bad decision to use a Kafka based stack. YMMV obviously.
That mixes the uses cases of analytics and operations because everyone is led to believe that things that happened in last 10 minutes must go through the analytics lens and yield actionable insights in real time so their operational systems can react/adapt instantly.
Most business processes probably don't need anywhere near such real time analytics capability but it is very easy to think (or be convinced that) we do. Especially if I am a owner of a given business process (with an IT budget) why wouldn't I want the ability to understand trends in real-time and react to it if not get ahead of them and predict/be prepared. Anything less than that is seen as being shamefully behind on the tech curve.
In this context-- the section in article where it says present data is of virtually zero importance to analytics is no longer true. We need a real solution even if we apply those (presumably complex and costly) solutions to only the most deserving use cases (and not abuse them).
What is the current thinking in this space? I am sure there are technical solutions here but what is the framework to evaluate which use case actually deserves pursuing such a setup.
Curious to hear.
https://www.snowflake.com/guides/htap-hybrid-transactional-a...
Anyone here taking them up on it? I'm genuinely curious how it's going.
> First, a working definition. An operational tool facilitates the day-to-day operation of your business. Think of it in contrast to analytical tools that facilitate historical analysis of your business to inform longer term resource allocation or strategy.
Security event and incident management (SEIM) is a typical example. You want fast notification on events combined with the ability to sift through history extremely quickly to assesss problems. This is precisely the niche occupied by real-time analytic databases like ClickHouse, Druid, and Pinot.
1. SNOW's market cap is $50B. GOOGL, MSFT, AMZN are all over $1T. Owning Snowflake would be a drop in the bucket for any of them (let alone if they were splitting the revenue).
2. Snowflake runs on AWS, GCP or Azure (customers choice), so a good chunk of their revenue goes back to these services.
Looking at these two points as the CEO of GOOGL, MSFT, or AMZN, I'd shrug away Snowflake "beating us". It's crazy that you can build a $50B company that your largest competitors barely care about.
BigQuery in GCP is a pretty great alternative and I know that GCP invests/promotes it heavily, but they were slightly late to the market.
FAANG can't utilize their market cap to buy SNOW, they would need to pay cash, and 50B is very large amount for any of these companies (its about annual Google net income).
Also, snow stock is very inflated now, it is heavily income negative, and revenue not that high, stock price is very high on growth expectations.
We aren't even going to consider the other direction? Running your analytics on top of a basic-ass SQL database?
In our shop, we aren't going for a separation between operational and analytical. The scale of our business and the technology available means we can use one big database for everything [0]. The only remaining challenge is to construct the schema such that consumers of the data are made aware of the rates of change and freshness of the rows (load interval, load date, etc).
If someone wants to join operational with analytical, I think they shouldn't have to reach for a weird abstraction. Just write SQL like you always would and be aware that certain data sources might change faster than others.
Sticking everything onto one target might sound like a risky thing, but I find many of these other "best practices" DW architectures to be far less palatable (aka sketchier) than one big box. Disaster recovery of 100% of our data is handled with replication of a single transaction log and is easy to validate.
[0]: https://learn.microsoft.com/en-us/azure/azure-sql/database/h...
Everytime I moved from one org to another, these concepts of data warehouse somehow got muddled.
is the usage of such an old html tag itself now a trigger to send something to /dev/null?
They indicate poorly written text.
One time I ran an A/B test on the color of a button. After the conclusion of the test, with a clear winner in hand, it took eleven months for all involved stakeholders to approve the change. The website in question got a few thousand visits a month and was not critical to any form of business.
This organization does not benefit from real-time analytics.
Now that's an extreme outlier, but my experience is that most organizations are in that position. The feedback loop from collecting data to making a decision is long, and real-time analytics shortens a part that's already not the bottleneck. The technical part of real-time analytics provides no value unless the org also has the operational capacity to use that data quickly.
I have seen this! I have, for example, seen a news site that looked at web analytics data from the morning and was able to publish new opinion pieces that afternoon if something was trending. They had a dedicated process built around that data pipeline. Critically, they had a specific idea of what they could do with that data when the received it.
So if you want a framework, I would start from a single, simple question: What can you actually do with real-time data? Name one (1) action your organization could take based on that data.
I think it's also useful to separate what data benefits from realtime and which users can make use of it. Even if you have real-time data, some consumers don't benefit from immediacy.
Real-time analytics are worse than useless. At best they are a distracting resource sink, at worst they directly harm the quality of decision-making.
Which is hard to explain because if it is not instant everywhere they think it is a bug and system is crappy. Later on they will use dashboard view once a week or once a month so 5 items update is not relevant at all.
I had this discussion with key people and i would say it depends on multiple factors. Small companies really like and require real-time analytics: they want to see how a couple invoices translate into updated saas metrics or why they didn’t get a slack/email notification as soon asit happened. Larger ones will check their data less frequently per day or week, but again it depends on the people and their role. Most of them are happy with getting their data once per day into their mailboxes or warehouses.
But we try to make everyone happy so we aim for real time analytics.
For most operations RT/NRT data stuff normally is about novelty/vanity rather than a real existing business need.
For some businesses, "real-time" can be properly defined as "within the last week". While there are many businesses where reducing that operational tempo to seconds would have an impact, it is by no means universal.
I will just point out that when my team and I talk about streaming, we are focused on not real-time because in many cases, the value to a customer is not there. Not every "streaming" use case is fraud detection. In fact, we have been saying for awhile that for many streaming use cases, the value is 60 seconds < [value here] < 60 minutes.
Example: (and yes, this is a Snowflake video but has a visual) https://youtu.be/Ou04UZWwxgg?t=64
On the other hand, real-time analytics for machines can be critical to a business, which is why Yandex built Clickhouse and ByteDance deployed more than 20K nodes of Clickhouse.
Just like using any technology, we need to figure out what problems we solve for real-time analytics first.
If asked people would say "I need to always be up" until they see the costs associated with it, then being out for a few hours a year tends to be ok.
There is a minimum cost though (systems, engineers, etc), so for medium data there's often very little marginal cost up until you start getting to hourly refreshes. This is not true for larger datasets though.
Totally agreed, though where real-time data is being put through an analytics lens is where CDW's start to creak and get costly. In my experience, these real-time uses shift the burden from being about human-decision-makers to automated decision-making and it becomes more a part of the product. And that's cool, but it gets costly, fast.
It also makes perfect sense to fake-it-til-you-make-it for real-time use cases on an existing Cloud Data Warehouse/dbt style _modern data stack_ if your data team's already using it for the rest of their data platform; after all they already know it and it's allowed that team to scale.
But a huge part of the challenge is that once you've made it, the alternative for a data-intensive use case is a bespoke microservice or a streaming pipeline, often in a language or on a platform that's foreign to the existing data team who's built the thing. If most of your code is dbt sql and airflow jobs, working with Kafka and streaming spark is pretty foreign (not to mention entirely outside of the observability infrastructure your team already has in place). Now we've got rewrites across languages/platforms, and leave teams with the cognitive overhead of multiple architectures & toolchains (and split focus). The alternative would be having a separate team to hand off real-time systems to and only that's if the company can afford to have that many engineers. Might as well just allocate that spend to your cloud budget and let the existing data team run up a crazy bill on Snowflake or BigQuery as long as it's less than the cost of a new engineering team.
------
There's something incredible about the ruthless efficiency of sql data platforms that allows data teams to scale the number of components/engineer. Once you have a Modern-Data-Stack system in place, the marginal cost of new pipelines or transformations is negligible (and they build atop one another). That platform-enabled compounding effect doesn't really occur with data-intensive microservices/streaming pipelines and means only the biggest business-critical applications (or skunk works shadow projects) will get the data-intensive applications[1] treatment, and business stakeholders will be hesitant to greenlight it.
I think Materialize is trying to build that Modern-Data-Stack type platform for real-time use cases: one that doesn't come with the cognitive cost of a completely separate architecture or the divide of completely separate teams and tools. If I already had a go-to system in place for streaming data that could be prototyped with the data warehouse, then shifted over to a streaming platform, the same teams could manage it and we'd actually get that cumulative compounding effect. Not to mention it becomes a lot easier to then justify using a real-time application the next time.
[1]: https://martin.kleppmann.com/2014/10/16/real-time-data-produ...
Building a database that can handle both analytics and operations is what we've been working on for the past 10+ years. Our customers use us to build applications with a strong analytical component to them (all of the use cases you mentioned and many more).
How's it going? It's going really well! And we're working on some really cool things that will expand our offering from being a pure data storage solution to much more of a platform[1].
If you want to learn more about our architecture, we published this paper at SIGMOD in late 2022 about it[2].
[1]: https://davidgomes.com/databases-cant-be-just-databases-anym...
Without some close to real world projections we don't have time to consider implementation to find out for ourselves.
Of course the number is going to be high, but you have to remember it rolls up compute and requires less manpower. This is also a win for finance if they are comfortable with usage based billing.
One of their most interesting offerings coming up is Snowpark which lets you run a Python function as a UDF, within Snowflake. This way you don't have to transfer data around everywhere, just run it as part of your normal SQL statements. It's also possible to pickle a function and send it over... so conceivably one could train a data science model and run that as part of a SQL statement. This could get very interesting.
https://www.snowflake.com/blog/snowpark-container-services-d...
Is that a differentiator? I'm unfamiliar with Snowpark's actual implementation but know SQL Server introduced Python/R in engine in 2016? something like that.
But in the end, snowflake stores the data in S3 as partitions. If you want to update a single value you have to replace the entire s3 partition. Similarly you need to read a reasonable amount of s3 data to retrieve even a single record. Thus you’re never going to get responses shorter than half a second (at best). As long as you don’t try and game around that limitation it works great.
Materialize up here also follows the same model in the end FWIW.
A common challenge in a lot of organizations is IT as a roadblock to deployment of internal tools coming from data teams. Snowflake is answering this with Streamlit. You get an easy platform for data people to use and deploy on and it can all be done within the business firewall under data governance within Snowflake.
Even with Vertica doing that, we're seeing 10x costs just doing back-office DWH. My job now is keeping Vertica running so we can pay our Snowflake bill.
I had frog getting boiled situation at one job that was exactly the process described in the posted article: usage of the cloud data warehouse grew as people trusted the infrastructure and used more fresh data for more and more use cases. They were all good, sane use cases. I repeatedly under-forecast our cost growth until we made large changes and it really frustrated the finance people, rightly so.
IIUC analytical = "what question are you trying to answer" and in analytics, RT/NRT is absolutely novelty/vanity. Operational = "what action are you trying to take" and it makes sense to want to have up-to-date data when, for example, running ML models, triggering notifications, etc...
(Also aren’t oracle the oracle of the cloud?)
Oh, you happened to run your oracle database in a VM and got audited? They'll try to shake you down to pay for however many oracle licenses for every other box you are running hypervisors on, because they claim that you could have transferred the database to any of those other boxes. So if you have a datacenter with 1000 boxes running VMWare, and you ran Oracle on one of them, they try to shake you down for paying for not buying 1000 licenses. Then they say but if you just buy a bunch of cloud credits, we can make your 1000x license violation go away.
People loved Oracle's technology back in the day (or so I understand it, I was like 2 when they released the first one).
It appears to make very different tradeoffs in a number of areas so that makes it a potential useful alternative. In particular transactional DML will make it much more convenient for workloads involving mutation. Plus as you suggested, having a proper Cost-Based Optimizer should make joins more efficient (I find ClickHouse joins to be fine for typical OLAP patterns but they do break down in creative queries...)
It's a bummer though that the deployment model is so complicated. One thing I truly like about ClickHouse is its ability to wring every drop of performance out of a single machine, with a super simple operational model. Being able to scale to a cluster is great but having to start there is Not Great.
Somehow StarRocks dudes appear in every relevant post with this false claim.
My experience with Clickhouse is that its joins are not performant enough to be useful. So the best practice in most cases is to denormalize. I should have been more specific in my earlier comment.
But that's fine! It has only existed for around a year in public preview, and appears to be improving quickly. My issue was with how aggressively Snowflake sales tried to push it as a production-ready ML platform. Whenever I asked questions about version control/CI, model versioning/ops, package managers, etc. the sales engineers and data scientists consistently oversold the product.
its common best practice on any database, because if both joined tables don't fit memory, then merge join is O(nlogn) operation which indeed many times slower than querying denormalized schema, which will have linear execution time.