Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Introducing dbt + Materialize

materialize.com

Post reply on HN
Print view View original

11–20 of 26 posts

Re: Introducing dbt + Materialize

#11

Post by jldlaughlin »

glogla wrote: post #10
jldlaughlin wrote: post #8

Earlier quoted context omitted.

We (I work at Materialize) actually do support updates! Our CDC sources support them, as well as any source using an UPSERT envelope (more info here: https://materialize.com/docs/sql/create-source/text-kafka/#u... ). As per your second point, I have a less awkward way for you! Materialize supports a top-k idiom ( https://materialize.com/docs/sql/idioms/#top-k-by-group ) that is hopefully a bit more clear.

Right, I saw no update in the SQL keyword list, and saw some (I guess old now) talk about Materialize where the speaker mentioned no updates, sorry!

No apologies necessary, we're an active work in progress! (And, hopefully, moving quickly!)

Re: Introducing dbt + Materialize

#12

Post by Nican »

benesch wrote: post #9
Nican wrote: post #4

Materialize is a really interesting solution, and I love for what it stands. But the documentation is missing more details about the architecture overview. A single update could cause many gigabytes of data to shift on the materialize view, and I do not understand how Materialize would handle that scale.

What you're asking about is the magic at the heart of Materialize. We're built atop an open-source incremental compute framework called Differential Dataflow [0] that one of our co-founders has been working on for ten years or so. The basic insight is that for many computations, when an update arrives, the amount of incremental compute that must be performed is tiny. If you're computing `SELECT count(1) FROM relation…

Thank you for taking your time to write that up. To illustrate my point a little more:

Suppose a customer has several resources, and each resource has several metrics. From my understanding, Materialized could be used to have an aggregated view of metrics per customer.

The problem is that resources can also be migrated between customers. When a resource migrates between customers, the whole history of the customer changes. This could cause huge updates depended on how many resources are moved, or how many metrics per resource are being collected.

I have a conundrum between doing the "customer-resource join" late, and causing huge CPU cost when running queries. Or making aggregates early, and then having huge Disk cost when migrating resources. At the moment, we just have daily jobs that aggregates the TBs of customer data daily, because there is no way to do the joins in real-time.

Is Materialize designed to be able to handle something like this?

Re: Introducing dbt + Materialize

#14

Post by albertwang »

d_watt wrote: post #5

Materialize seems really, really interesting. To the point where my crotchety old man senses are telling me that I should be careful about getting too excited about new technology. Are there any interesting case studies of people pushing the edge of what's possible? For my use case, I have max throughput of 10k updates/second potentially materialized into tens of thousands of different views (with some good partition…

Nothing we can share publicly at the moment yet, but if you reach out and chat, we're more than happy to give you some numbers that I think will address what you're looking for!

Re: Introducing dbt + Materialize

#15

Post by jdotjdot »

I have been waiting for this since the moment I first read about Materialize a year or two ago. I think there's still a lot of work to be done, but at heart, if you can pair technology like Materialize with an orchestration system like dbt, you can use dbt to keep your business logic extremely well organized, yet have all of your dependent views up to date all of the time, and use dbt even to use the same analytical layered views both for analytical AND operational purposes.

The biggest issue I see is that it requires you to be all-in on Materialize, and as a warehouse (or as a database for that matter), it's surely not as mature as Snowflake or Postgres.

Re: Introducing dbt + Materialize

#16

Post by jldlaughlin »

efangs wrote: post #13

Is it possible to backfill a materialized view? I'm maybe confused, but does this potentially replace a traditional data lake or is it specifically for streaming applications?

It is! We recently added support for S3 sources [0], which you could use to backfill data and union with a stream.

To your other question, we're currently well-suited for streaming applications. Moving forward, as we add support for features like persistence, we could certainly replace at least parts of a traditional data lake.

[0]: https://materialize.com/docs/sql/create-source/json-s3/

Re: Introducing dbt + Materialize

#17

Post by arjunnarayan »

jdotjdot wrote: post #15

I have been waiting for this since the moment I first read about Materialize a year or two ago. I think there's still a lot of work to be done, but at heart, if you can pair technology like Materialize with an orchestration system like dbt, you can use dbt to keep your business logic extremely well organized, yet have all of your dependent views up to date all of the time, and use dbt even to use the same analytical…

Thank you for your kind words! We indeed have plenty of work to be done (and are thus hiring)! I'm curious however why you think this requires you to be all-in on Materialize. As you said better than I could have, dbt is amazing at keeping your business logic organized. Our intention is very much for dbt to standardize the modeling/business logic layer which allows you to use multiple backends as you see fit in a way that shares the catalog layer cleanly.

Our hope is that you have some BigQuery/Snowflake job that you're tired of running up the bill hitting redeploy 5 times a day, and you can cleanly port that over to Materialize with little work because the adapter is taking care of any small semantic differences in date handling, or null handling, etc. So Materialize sits cleanly side-by-side with Snowflake/BigQuery, and you're choosing whether you want things incrementally maintained with a few seconds of latency by Materialize, or once a day by the batch systems.

My view is you're likely going to want to do data science with a batch system (when you're in "learning mode" you try and keep as many things fixed, including not updating the dataset), and then if the model becomes a critical automated pipeline, rather than rerunning the model every hour and uploading results to a Redis cache or something, you switch it over to Materialize, and don't have to every worry about cache invalidation.

Re: Introducing dbt + Materialize

#18

Post by benesch »

Nican wrote: post #12
benesch wrote: post #9

Earlier quoted context omitted.

What you're asking about is the magic at the heart of Materialize. We're built atop an open-source incremental compute framework called Differential Dataflow [0] that one of our co-founders has been working on for ten years or so. The basic insight is that for many computations, when an update arrives, the amount of incremental compute that must be performed is tiny. If you're computing `SELECT count(1) FROM relation…

Thank you for taking your time to write that up. To illustrate my point a little more: Suppose a customer has several resources, and each resource has several metrics. From my understanding, Materialized could be used to have an aggregated view of metrics per customer. The problem is that resources can also be migrated between customers. When a resource migrates between customers, the whole history of the customer ch…

So, roughly speaking, I think you're describing a view that could be defined in Materialize like so:

    CREATE MATERIALIZED VIEW customer_avg_latency
    SELECT customer_id, metric_id, avg(metric_val)
    FROM metrics
    JOIN resources ON metrics.resource_id = resources.id
    JOIN customers ON resources.customer_id = customer.id
    GROUP BY customer_id, metric_id
There are various ways to slice and dice that query as a user that'll allow you to choose whether to do the aggregate early or late. The details, I think, depend on the specifics of your data model.

> When a resource migrates between customers, the whole history of the customer changes.

I'm a bit confused about this part. In the query I posted above, Differential's incremental join algorithm would spare you from recomputing the entire join. (As written the query would have to re-aggregate the metrics associated with that resource that moved, but that could be solved by manually pushing the avg down.)

If you'd like to drop by our Community Slack (https://materialize.com/s/chat) or GitHub issues (https://github.com/MaterializeInc/materialize/issues) we'd be happy to talk through things in more detail!

Re: Introducing dbt + Materialize

#19

Post by abrazensunset »

arjunnarayan wrote: post #17
jdotjdot wrote: post #15

I have been waiting for this since the moment I first read about Materialize a year or two ago. I think there's still a lot of work to be done, but at heart, if you can pair technology like Materialize with an orchestration system like dbt, you can use dbt to keep your business logic extremely well organized, yet have all of your dependent views up to date all of the time, and use dbt even to use the same analytical…

Thank you for your kind words! We indeed have plenty of work to be done (and are thus hiring)! I'm curious however why you think this requires you to be all-in on Materialize. As you said better than I could have, dbt is amazing at keeping your business logic organized. Our intention is very much for dbt to standardize the modeling/business logic layer which allows you to use multiple backends as you see fit in a way…

In that situation (dual usage modes) I think I'd rather have the primary data store be Materialize, and just snapshot Materialize views back to your warehouse (or even just to an object store).

Then you could use that static store for exploration/fixed analysis or even initial development of dbt models for the Materialize layer, using the Snowflake or Spark connectors at first. When something's ready for production use, migrate it to your Materialize dbt project.

The way dbt currently works with backend switching (and the divergence of SQL dialects with respect to things like date functions and unstructured data), maintaining the batch and streaming layers side by side in dbt would be less wasteful than the current paradigm of completely separate tooling, but still a big source of overhead and synchronization errors.

If the community comes up with a good narrative for CI/CD and data testing in flight with the above, I don't think I'd even hesitate to pull the trigger on a migration. The best part is half of your potential customers already have their business logic in dbt.

Post reply on HN