650GB of Data (Delta Lake on S3). Polars vs. DuckDB vs. Daft vs. Spark

https://dataengineeringcentral.substack.com/p/650gb-of-data-delta-lake-on-s3-polars

by tanelpoder • 8 months ago

263 107 8 months ago

dataengineeringcentral.substack.com

esafak3• 8 months ago

If I understand correctly, polars relies on delta-rs for Delta Lake support, and that is what does not support Deletion vectors: https://github.com/delta-io/delta-rs/issues/1094

It seems like these single-node libraries can process a terabyte on a typical machine, and you'd have have over 10TB before moving to Spark.

mynameisash• 8 months ago

It seems like these single-node libraries can process a terabyte on a typical machine, and you'd have have over 10TB before moving to Spark.

I'm surprised by how often people jump to Spark because "it's (highly) parallelizable!" and "you can throw more nodes at it easy-peasy!" And yet, there are so many cases where you can just do things with better tools.

Like the time a junior engineer asked for help processing 100s of ~5GB files of JSON data which turned out to be doing crazy amounts of string concatenation in Python (don't ask). It was taking something like 18 hours to run, IIRC, and writing a simple console tool to do the heavy lifting and letting Python's multiprocessing tackle it dropped the time to like 35 minutes.

Right cool for the right job, people.

esafak3• 8 months ago

I used pySpark some time ago when it was introduced to my company at the time and I realized that it was slow when you used python libraries in the UDFs rather than pySpark's own functions.

rgblambda• 8 months ago

I think Spark was the best tool out there when data engineering started taking off, and it just works (provided you don't have to deal with jar dependency hell) so there's not a huge incentive to move away from it.

nijave2• 8 months ago

Python isn't too bad if you swap in orjson instead of stdlib which is quite a bit slower

Wrangling multiprocess is still annoying tho

blmarket• 8 months ago

Presto (a.k.a. AWS Athena) might be a faster/better alternative? Also would like to see if 650GB data is available locally.

fifilura• 8 months ago

Presto is renamed to Trino now.

But I concur with what you say. It is also very cheap in both maintenance and running cost. It is just an amazing tool and you pay (RIP) pennies.

tbcj• 8 months ago

No, Presto (https://github.com/prestodb/presto) remains alive and well, just Trino gets more attention.

andy99• 8 months ago

Awk? https://adamdrake.com/command-line-tools-can-be-235x-faster-...

co0lster• 8 months ago

650GB relates to size of parquet files which are compressed in reality it’s way more.

32 GB of parquet cannot fit in 32GB of RAM

barrkel• 8 months ago

This would speed things up since it looks like the bottleneck here is I/O.

m00x• 8 months ago

You don't need it to if you just need specific columns. This is the advantage of columnar storage.

luizfelberti2• 8 months ago

Honestly this benchmark feels completely dominated by the instance's NIC capacity.

They used a c5.4xlarge that has peak 10Gbps bandwidth, which at a constant 100% saturation would take in the ballpark of 9 minutes to load those 650GB from S3, making those 9 minutes your best case scenario for pulling the data (without even considering writing it back!)

Minute differences in how these query engines schedule IO would have drastic effects in the benchmark outcomes, and I doubt the query engine itself was constantly fed during this workload, especially when evaluating DuckDB and Polars.

The irony of workloads like this is that it might be cheaper to pay for a gigantic instance to run the query and finish it quicker, than to pay for a cheaper instance taking several times longer.

amluto2• 8 months ago

It would be amusing to run this on a regular desktop computer or even a moderately nice laptop (with a fan - give it a chance!) and see how it does. 650GB will stream in quite quickly from any decent NVMe device, and those 8-16 cores might well be considerably faster than whatever cores the cloud machines are giving you.

S3 is an amazingly engineered product, operates at truly impressive scale, is quite reasonably priced if you think of it as warm-to-very-cold storage with excellent durability properties, and has performance that barely holds a candle to any decent modern local storage device.

tempest_• 8 months ago

Everyone wants a data lake when what they have a is a data pond.

switchbak• 8 months ago

Absolutely. I recently reworked a bunch of tests and found my desktop to outcompete our (larger, custom) Github Action runner by roughly 5x. And I expect this delta to increase a lot as you lean on the local I/O harder.

It really is shocking how much you're paying given how little you get. I certainly don't want to run a data center and handle all the scaling and complexity of such an endeavour. But wow, the tax you pay to have someone manage all that is staggering.

layoric• 8 months ago

Totally true. I have a trusty old (like 2016 era) X99 setup that I use for 1.2TB of time series data hosted in a timescaledb PostGIS database. I can fetch all the data I need quickly to crunch on another local machine, and max out my aging network gear to experiment with different model training scenarios. It cost me ~$500 to build the machine, and it stays off when I'm not using it.

Much easier obviously dealing with a dataset that doesn't change, but doing the same in the cloud would just be throwing money away.

bushbaba• 8 months ago

I'm kind of suprised they didn't choose an ec2 instance with higher throughput. S3 can totally eek out 100s of Gibps with the right setup.

BUT the author did say this is the simple stupid naive take, in which case DuckDB and Polars really shined.

kccqzy2• 8 months ago

10Gbps only? At Google where this type of processing would automatically be distributed, machines had 400Gbps NICs, not to mention other innovations like better TCP congestion control algorithms. No wonder people are tired of distributed computing.

basilgohar• 8 months ago

"At Google" is doing all the heavy lifting in your comment here, with all due respect. There is but one Google but remain millions of us who are not "At Google".

otterley• 8 months ago

You can get a 600Gbps interface on an Amazon EC2 instance (c8gn.48xlarge), if you’re willing to pay for it.

nijave2• 8 months ago

They do at AWS, too but op paid for a small VM

dukodk• 8 months ago

c5 is such a bad instance type, m6a would be so much better and even cheaper, I would love to see this on an m8a.2xlarge (7th and 8th generations don’t use SMT) and that is even cheaper and has up to 15 Gbps

luizfelberti2• 8 months ago

Actually for this kind of workload 15Gbps is still mediocre. What you actually want is the `n` variant of the instance types, which have higher NIC capacity.

In the c6n and m6n and maybe the upper-end 5th gens you can get 100Gbps NICs, and if you look at the 8th gen instances like the c8gn family, you can even get instances with 600Gbps of bandwidth.

mrbungie• 8 months ago

They used a c5.4xlarge that has peak 10Gbps bandwidth, which at a constant 100% saturation would take in the ballpark of 9 minutes to load those 650GB from S3, making those 9 minutes your best case scenario for pulling the data (without even considering writing it back!)

The query being tested wouldn't scan the full files and in reality the query in most sane engines would be processing much less than 650GB of data (exploiting S3 byte-range reads): i.e. just 1 column: a timestamp, which is also correlated with the partition keys. Nowadays what I would mostly be worried about the distribution of file size, due to API calls + skew; or if the query is totally different to the common query access patterns that skips the metadata/columnar nature of the underlying parquet (i.e. doing an effective "full scan" over all row groups and/or columns).

The irony of workloads like this is that it might be cheaper to pay for a gigantic instance to run the query and finish it quicker, than to pay for a cheaper instance taking several times longer.

That's absolutely right.

mrlongroots3• 8 months ago

Yep I think the value of the experiment is not clear.

You want to use Spark for a large dataset with multiple stages. In this case, their I/O bandwidth is 1GB/s from S3. CPU memory bandwidth is 100-200GB/s for a multi-stage job. Spark is a way to pool memory for a large dataset with multiple stages, and use cluster-internal network bandwidth to do shuffling instead of storage.

Maybe when you have S3 as your backend, the storage bandwidth bottleneck doesn't show up in perf, but it sure does show up in the bill. A crude rule of thumb: network bandwidth is 20X storage, main memory bandwidth is 20X network bandwidth, accelerator/GPU memory is 10X CPU. It's great that single-node DuckDB/Polars are that good, but this is like racing a taxiing aircraft against motorbikes.

justincormack• 8 months ago

Network bandwidth is not 20x storage ant more. An SSD is around 10GB/s now, so similar to 100Gb ethernet.

Scubabear68• 8 months ago

This is a really good observation, and matches something I had to learn painfully over 30 years ago. At a Wall Street bank, we were trying to really push the limits with some middleware, and my mentor at the time very quietly suggested "before you test your system's performance, understand the theoretical maximum of your setup first with no work".

The gist was - find your resource limits and saturate them and see what the best possible performance could be, then measure your system, and you can express it as a percentage of optimal. Or if you can't directly test/saturate your limits at least be aware of them.

jdnier• 8 months ago

DuckDb has a new "DuckLake" catalog format that would be another candidate to test. https://ducklake.select/

garganzol• 8 months ago

DuckLake format has an unresolved built-in chicken and egg conflict: it requires SQL database to represent its catalog. But this is what some people are running away from when they choose Parquet format in the first place. Parquet = easy, SQL = hard, adding SQL to Parquet makes the resulting format hard. I would expect a catalog to be in Parquet format as well, then it becomes something self-bootstrapping and usable.

matt123456789• 8 months ago

It is not a chicken and egg problem, it is just a requirement to have an RDBMS available for systems like DuckLake and Hive to store their catalogs in. Metadata is relatively small and needs to provide ACID r/w => great RDBMS use case.

datacynic• 8 months ago

DuckLake is more comparable to Iceberg and Delta than to raw parquet files. Iceberg requires a catalog layer too, a file system based one at its simplest. For DuckLake any RDBMS will do, including fs-based ones like DuckDB and SQLite. The difference is that DuckLake will use that database with all its ACID goodness for all metadata operations and there is no need to implement transactional semantics over a REST or object storage API.

sukhavati• 8 months ago

for me the issue is that DuckLake's feature of flushing inlined data to parquet is still in alpha. one of the main issues with parquet is when writing small batches you end up with a lot of parquet files that are inefficient to work with using duckdb. to solve this ducklake inlines these small writes to the dbms you choose (postgres) but for a while it couldn't write them back to parquet. last I had checked this feature didn't yet exist, and now it seems to be in alpha which is nice to see, but I'd like some better support before I consider switching some personal data projects over. https://ducklake.select/docs/stable/duckdb/advanced_features...

erikcw• 8 months ago

Data inlining is also currently limited to only the DuckDB catalog (ie it doesn't work with Postgres cataglogs)[0]. It's improving very quickly though and I'm sure this will be expanded soon.

[0] https://ducklake.select/docs/stable/duckdb/advanced_features...

abofh• 8 months ago

6$ of data does not a compelling story make. This is not 1998

hnidiots3• 8 months ago

650GB? We have 72PB IN S3, know people who have multiple EB in S3.

0cf8612b2e1e2• 8 months ago

This is not a game of, “mine is bigger than yours”. Many many workloads in the wild are smaller than this.

Motherduck have a few posts about how few people have “big data”. https://motherduck.com/blog/redshift-files-hunt-for-big-data...

esafak3• 8 months ago

That has to be multiple data sets. How big are your individual nightly jobs, and what are you processing them with?

gdevenyi• 8 months ago

I hate this screenshots for commands and outputs everywhere

tacker2000• 8 months ago

I also hate it. Cant read it properly on mobile and cant copy it if needed.

But the worst thing is that these are not even real screenshots, the author pasted the text into some terminal window screenshot generator tool.

dogman123• 8 months ago

One thing that I never really see mentioned in these types of articles is that a lot of DuckDB’s functionality does not work if you need to spill to disk. iirc, percentiles/quartiles (among other aggregate functions) caused DuckDB to crash out when it spilled to disk.

jtbaker• 8 months ago

I’m pretty sure I’ve done this and not had any issues. Can you share a minimum reproducible example?

srilman• 8 months ago

Hey everyone, I'm a software engineer at Eventual, the team behind Daft! Huge thanks to the op for the benchmark, we're a huge fan of your blog posts and this gave us some really useful insights. For context, Daft is a high-performance data processing engine for AI workloads that works both on single-node and distributed setups.

We're actively looking into the results of the benchmark and hope to share some of our findings soon. From initial results, we found a lot of potential optimizations we could make to our deltalake reader to improve parallelism and our groupby operator to improve pipelining for count aggregations. We're hoping to roll our these improvements over the next couple of releases.

If you're interested to learn more about our findings, check out our GitHub (https://github.com/Eventual-Inc/Daft) or follow us on Twitter (https://x.com/daftengine) and LinkedIn (https://www.linkedin.com/showcase/daftengine) for updates. Also if Daft sounds interesting to you, give us a try via pip install daft!

CyberDildonics• 8 months ago

This looks like you made an account to post one comment to advertise your company.

hersko• 8 months ago

So?

VHRanger• 8 months ago

Do you plan to expose daft as a backend in ibis?

That would be the best way to smoothly test it out and transition workloads from other engines for codebases in my teams.

throwaway-aws9• 8 months ago

650GB? Your data is small, fits on my phone. Dump the hyped tooling and just use gnu tools.

Here's an oldie on the topic: https://adamdrake.com/command-line-tools-can-be-235x-faster-...

Demiurge2• 8 months ago

What if it was 650TB? This article is obviously a microbenchmark. I work with much larger datasets, and neither awk nor DBD would make a difference to the overall architecture. You need a data catalog, and you need a clusters of jobs at scale, regardless of a data format library, or libraries.

CraigJPerry• 8 months ago

At 650tb it's not a memory bound problem:

working memory requirements

    1. Assume date is 8 bytes
    2. Assume 64bit counters

So for each date in the dataset we need 16 bytes to accumulate the result.

That's ~180 years worth of daily post counts per gb ram - but the dataset in the post was just 1 year.

This problem should be mostly network limited in the OP's context, decompressing snappy compressed parquet should be circa 1gb/sec. The "work" of parsing a string to a date and accumulating isn't expensive compared to snappy decompression.

I don't have a handle on the 33% longer runtime difference between duckdb and polars here.

adammarples• 8 months ago

I think the entire point of the article (reading forward a bit through the linked redshift files posts) is that almost nobody in the world uses datasets bigger than 100Tb, that when they do, they use a small subset anyway, and that 650Gb is a pretty reasonable approximation of the entire dataset most companies are even working with. Certainly in my experience as a data engineer, they're not often in the many terabytes. It's good to know that OOTB duckdb can replace snowflake et all in these situations, especially with how expensive they are.

jgalt212• 8 months ago

"I've forgotten how to count that low"

https://www.youtube.com/watch?v=3t6L-FlfeaI

faizshah2• 8 months ago

This isn’t true anymore we are way beyond 2014 Hadoop (what the blog post is about) at this point.

Go try doing an aggregation of 650gb of json data using normal CLI tools vs duckdb or clickhouse. These tools are pipelining and parallelizing in a way that isn’t easy to do with just GNU Parallel (trust me, I’ve tried).

willvarfar2• 8 months ago

I often crunch 'biggish data' on a single node using duckdb (because I love using the modern style of painless and efficient SQL engines).

I don't use delta or iceberg (because I haven't needed to; I'm describing what I do, not what you can do :)), but rather just iterate over the underlying parquet files using filename listing or wildcarding. I often run queries on BigQuery and suck down the results to a bunch of ~1GB local parquet files - way bigger than RAM - that I can then mine in duckdb using wildcarding. Works great!

I'm in a world where I get into the weeds of 'this kind of aggregation works much faster on Bigquery than duckdb, or vice versa, so I'll split my job into this part of sql running on Bigquery then feeding into this part running in duckdb'. It's the fun end of data engineering.

miohtama• 8 months ago

650GB is something one could handle using a local filesystem, no need for complex tooling.

capitol_• 8 months ago

650 GB fits in ram: https://yourdatafitsinram.net

nevi-me• 8 months ago

The main reason why clusters still make sense is because you'll have a bunch of people accessing subsets of much larger data regularly, or competing processes that need to have their output ready at around the same time. You distribute not only compute, but also I/O, which others are pointing out to likely dominate the runtime of the benchmarks.

Beyond Spark (one shouldn't really be using vanilla Spark anyways, see Apache Comet or Databricks Photon), distributing my compute makes sense because if a job takes an hour to run, (ignoring overnight jobs) there will be a bunch of people waiting for that data for an hour.

If I run a 6 node cluster that makes the data available in 10 minutes, then I save in waiting time. And if I have 10 of those jobs that need to run at the same time, then I need a burst of compute to handle that.

That 6 node cluster might not make sense on-prem unless I can use the compute for something else, which is where PAYG on some cloud vendor makes sense.

jiehong• 8 months ago

This is somewhat real world, except real world would probably index some parquet columns to avoid a full scan like that.

zkmon2• 8 months ago

There are other factors as well, that drive the decision makers to clusters and big-data tech, even when the benchmarks do not justify that. At the root, the reasons are organizational, not technical. Risk aversion seeks to avoid single point of failure, needs accountability, favors outsourcing to specialists etc. Performance alone is not going to beat all of that.

willvarfar2• 8 months ago

Often, at the medium and large sized companies its not 'risk aversion', its resume padding.

Architects want to build big impressive systems that justify their position and managers want that too because success is judged by size of systems and number of staff under management, not its efficiency; its all about perverse incentives.

This is just a tax the scientists trying to use whatever the company settles on have to pay every time they wait for queries to run.

These days scientists can just suck down a copy of a bunch of data to their laptop or a cheap cloud VM and do their crunching 'locally' there. The company data swamp is just something they have to interface with occasionally.

Of course things go pear-shaped if they get detected, so don't tell anyone :D

zkmon2• 8 months ago

Quite true. There are hardly any technical justifications for this madness, other than seeking a bloat of work and team size at the expense of huge spend.

pu_pe• 8 months ago

In places I have worked at that used Databricks, I feel they chose it for the same reasons big orgs use Microsoft: it comes out of a box and has a big company behind it. Technical benchmarks or even cost considerations would be a distant second.

data_marsupial• 8 months ago

There are real advantages from having a managed data platform compared to managing everything yourself, especially if you have a large number of data teams that need to collaborate.

hobs• 8 months ago

Yep, and Databricks will have you churning and changing everything on your stack every 18 months (if you want to keep up to date at all) - its not what I would choose as a data partner unless I was just picking what all the other kids at lunch were.

rmnclmnt2• 8 months ago

Until the product manager ask for the bill… then all of a sudden things get reconsidered

patwolf• 8 months ago

I worked at a company that dumped Databricks once the first bill came. I guess it was an order of magnitude more expensive than what they expected. It was less expensive to rebuild the pipeline from scratch with a different product.

jellyfishbeaver2• 8 months ago

Are there really any other managed Spark/full data platform providers that come close to the level of ergonomics and maturity of Databricks? I agree some of their features are half-baked and you can turn on some dangerous cash faucets if you don't have a very diligent administrator, but I have not come across anything that our developers (fresh grads up to seasoned Spark veterans) liked as much as Databricks.

baq2• 8 months ago

650GB would’ve fit in a not-exotic-at-all basically off the shelf server ram a decade ago

mettamage• 8 months ago

Truly, we have not been thinking outside the box with the Modern Lake House architecture. Just because Pandas failed us doesn’t mean distributed computing is our only option.

Well yea, I would have picked polars as well. To be fair , I didn’t know about some of these.

nikita2206• 8 months ago

I am not in data eng, but I do occasionally query data lake at my company. Where does Snowflake stand in this? (specially looking at that Modern Data Stack image)

benrutter3• 8 months ago

I beleive snowflake has its own distributed query engine, similar to say, big query.

It's a bit of a tricky comparison because snowflake, and a lot of other tools that get referred to as "modern data stack" are very vendor based. If you're using snowflake, you're probaby using it on snowflake provided architecture with a whole load of proprietary stuff. You can't "swap in" snowflake on the same hardware like you can with spark, daft, duckdb, polars etc.

That said, iirc benchmarks normally place it very similar to spark. It's distributed, so I'd be very surprised if it wasn't in the spark/daft ballpark rather than polars/duckdb.

throwaw12• 8 months ago

I am curious as well about this, we use Snowflake, but as a software engineer I want to understand how Spark/Databricks is different, what are we missing out?

How we work with data is simple, if SQL+dashboard solves the problem then we do it in Snowflake, if we need something more advanced, then code + bunch of SQL.

Pretty sure ML engineers work in different ways, but I don't know that side well

roeja• 8 months ago

Snowflake has their own sql engine and is more of a serverless option. Databricks started off with spark but now also has a sql engine(optional serverless) as well, they are using spark in the article.

The delta format is Databricks lakehouse file format, snowflake uses iceberg I believe.

Both Snowflake and Databricks also provide a ton of other features like ML, Orchestration and governance. Motherduck would be the direct competitor here.

Saying that there are now extensions to query snowflake or databricks data from duckdb for simple ad hoc querying.

Duckdb is fantastic and has saved me so many times strongly recommended.

prpl• 8 months ago

"but now also has a sql engine"

it has had a combined SQL and dataframe engine since March 2015...

benrutter3• 8 months ago

I love this article! But I think this insight shouldn't be surprising. Distribution always has overheads, so if you can do things on a single machine it will almost always be faster.

I think a lot of engineers expect 100 computers to be faster than 1, because of the size comparison. But we're really looking at a process here, and a process shifting data between machines will almost always have to do more stuff, and therefore be slower.

Where spark/daft are needed is if you have 1tb of data or something crazy were a single machine isn't viable. If I'm honest though, I've seen a lot of occasions where someone thinks they have that happening, and none so far where they actually do.

0cf8612b2e1e2• 8 months ago

The Scalability at what COST paper (pdf https://www.usenix.org/system/files/conference/hotos15/hotos...) is my favorite thing. Single worker implementation wipes the floor with big distributed solutions.

zigzag312• 8 months ago

DataFusion is another option I would be interested to see in a comparison like this.

ayhanfuat2• 8 months ago

What is the point of simulating 650GB data with ~40 columns if you are going to use a single column for testing? Is that even 16GB?

KeplerBoy3• 8 months ago

It's a strided array and slows down memory access.

ayhanfuat2• 8 months ago

It's a parquet file. Column data is stored in contiguous pages (and that's how duckdb and polars read them).

l_c_m• 8 months ago

This is most misrepresented article on two fronts

1. tested column pruning and the dataset you access would have been 2 columns + metadata for the parquet files so probably fit in memory even without streaming.

2. Most of the processing time would be IO bound on S3 and the access patterns/simultaneous connection limits etc. would have more of an impact than any processing code.

Love that you went through the pain of trying the different systems but I'd like to see an actual larger than memory query.

jamesblonde• 8 months ago

1. Important points that the query is a projection that only returns a fraction of the 650GB that fits in memory. DuckDB is good at streaming larger than memory queries, Polars less mature there. That would show in the results.

2. S3 defaults shouldn't prevent all available threads/cpus from reading the files in parallel, so I would assume that the network bandwidth of the VM (or container) would be the bottleneck.

DiskoHexyl• 8 months ago

Hardly a surprise, given the nature of Spark and benchmark prerequisites. Comparing a positively ancient distributed JVM-based compute framework running on a single node, with modern native tools like DuckDB or Polars, and all that on a select from a single table- does it tell us something new?

Even Trino runs circles around Spark, with some heavier jobs simply not completing in Spark at all (total data size up to a single PB, with about 10TB of RAM available for compute), and Trino isn't known for its extreme performance. StarRocks is noticeably faster still, so I wouldn't right off distributed compute just yet- at least for some applications.

And even then, performance isn't the most important criterion for an analytics tool choice- more probably depends on the integrations, access control, security, ease of extendability, maintenance, scaling, support by existing instruments. Boring enterprise stuff, sure, but for those older frameworks it's all either readily available, or can be quickly added with little experience (writing a java plugin for Trino is as easy as it gets).

With Duckdb or Polars (if used as a basis for a datalake/house etc) it may degrade into an entire team of engineers wasting resources on implementing the tooling around the tooling instead of providing something actually useful for the business

faizshah2• 8 months ago

I had to do something like this for a few TB of json recently. The unique thing about this workload was it was a ton of small 10-20mb files.

I found that clickhouse was the fastest, but duckdb was the simplest to work with it usually just works. DuckDB was close enough to the max performance from clickhouse.

I tried flink & pyspark but they were way slower (like 3-5x) than clickhouse and the code was kind of annoying. Dask and Ray were also way too slow, but dask’s parallelism was easy to code but it was just too slow. I also tried Datafusion and polars but clickhouse ended up being faster.

These days I would recommend starting with DuckDB or Clickhouse for most workloads just cause it’s the easiest to work with AND has good performance. Personally I switched to using DuckDB instead of polars for most things where pandas is too slow.

sagarm2• 8 months ago

Did you first ingest/convert this data to some other format, or did you operate directly on the JSON?

knorker• 8 months ago

Your data fits in RAM.

If this is a test of how well the process would work for big data sets, then I'm not sure it's even big enough to be able to extrapolate from.

arconis987• 8 months ago

would like to see work like this, but for datasets in the hundreds of TB or single-digit PB

but i definitely agree about this point

Cluster fatigue is real

imo, the concept of “extremely ephemeral query workers” is under-explored

stateless, maintenance-free, burstable fleets of query workers is what I would like to see more of in the future.

it’s how we do it, and it gives us full-text search on multi-hundred terabyte data sets in S3, where queries finish in a handful of seconds. our approach: https://docs.scanner.dev/scanner/what-and-why/how-it-works/h...

anyone else doing ephemeral query workers fleets?

ramraj07• 8 months ago

Yes.. its called snowflake? Theyre exactly that and why they work so well. I know youre asking for an OSS but what snowflake offers is a fleet of servers that can build your cluster in a second as opposed to minutes that you need if you want to spin it up yourself..

sagarm2• 8 months ago

extremely ephemeral query workers

Reading data from S3 can really add up, so this isn't as straightforward as it seems.

ardmiller• 8 months ago

Not super related but one complaint about Polars is I don't get why the departure from the pandas API...

cloud-oak• 8 months ago

I was very reluctant about the polars syntax as well initally, but it has grown a lot on me.

Pandas syntax is super ergonomic for quick one-off analysis, but it becomes hard to read/maintain once your processing gets more complex.

For example, the innocent

    df[arg]

can mean wildly different things - does it filter rows? Subset columns? Extract a single column as a pd.Series? There really is no way of knowing except for checking the value of arg.

In contrast, polars syntax feals clunky initially, but it's much easier for me to revisit a pipeline and quickly understand what it does.

pancakeguy• 8 months ago

It's cool that this is possible on a single node but I still think distributed is the way.

The point of these tools is productivity. What are you trying to accomplish and how long does it take to accomplish? This includes time spent writing code and fussing with configs. This would take <1min to run and <10 to write on a cluster. Happy to might make a demo to prove it.

Yes cost matters also, but running many machines for a short period of time is the same as one for a long time? Open to honest rebuttal.

650GB of Data (Delta Lake on S3). Polars vs. DuckDB vs. Daft vs. Spark

Related Stories