DuckLake is an integrated data lake and catalog format
94 comments
·May 27, 2025amluto
bitbang
Why is the footer metadata not sufficient for this need? The metadata should contain the min and max timestamp values from the respective column of interest, so that when executing a query, the query tool should be optimizing its query by reading the metadata to determine if that parquet file should be read or not depending on what time range is in the query.
amluto
Because the footer metadata is in the Parquet file, which is already far too late to give an efficient query.
If I have an S3 bucket containing five years worth of Parquet files, each covering a few days worth of rows, and I tell my favorite query tool (DuckDB, etc) about that bucket, then the tool will need to do a partial read (which is multiple operations, I think, since it will need to find the footer and then read the footer) of ~500 files just to find out which ones contain the data of interest. A good query plan would be to do a single list operation on the bucket to find the file names and then to read the file or files needed to answer my query.
Iceberg and Delta Lake (I think -- I haven't actually tried it) can do this, but plain Parquet plus Hive partitioning can't, and I'm not aware of any other lightweight scheme that is well supported that can do it. My personal little query tool (which predates Parquet) can do it just fine by the simple expedient of reading directory names.
dugmartin
This can also be done using row group metadata within the parquet file. The row group metadata can include the range values of ordinals so you can "partition" on timestamps without having to have a file per time range.
amluto
But I want a file per range! I’m already writing out an entire chunk of rows, and that chunk is a good size for a Parquet file, and that chunk doesn’t overlap the previous chunk.
Sure, metadata in the Parquet file handles this, but a query planner has to read that metadata, whereas a sensible way to stick the metadata in the file path would allow avoiding reading the file at all.
simlevesque
I wish we had more control of the row group metadata when writing Parquet files with DuckDB.
null
TheCondor
Hive supports 2 kinds of partitioning, injected and dynamic. You can totally use a partition key like the hour in UNIX time. It's an integer starting at some epoch and incrementing by 3600.
Now your query engine might require you to specify the partitions or range of partitions you want to query on; you absolutely can use datepartition >=a and datepartition<b in your query. Iceberg seems to fix that and just let you use the timestamp; presumably the metadata is smart enough to exclude the partitions you don't care about.
amluto
This is exactly what I meant by “I could also partition on a silly column like "date". This involves aligning batches to date boundaries and also makes queries uglier.”
hendiatris
In the lower level arrow/parquet libraries you can control the row groups, and even the data pages (although it’s a lot more work). I have used this heavily with the arrow-rs crate to drastically improve (like 10x) how quickly data could be queried from files. Some row groups will have just a few rows, others will have thousands, but being able to bypass searching in many row groups makes the skew irrelevant.
Just beware that one issue you can have is the limit of row groups per file (2^15).
jonstewart
I think maybe this is a pet peeve of Hive and not of Parquet? Yes, it does require opening the Parquet file to look at the min, max range for the column, but only that data and if the data isn’t in range there shouldn’t be further requests.
That is the kind of metadata that is useful to push up, into something like DuckLake.
amluto
I guess my peeve could be restated as: Hive’s naming scheme doesn’t handle this, and Parquet per se can’t handle it because it’s out of scope, but all the awesome tools (e.g. DuckDB), when used on Parquet files without something with “lake” or “ice” in the name, use the Hive scheme.
Someone could buck the trend and extend Hive’s scheme to support ranges.
dkdcio
This looks awesome. One of my biggest gripe's personally with Iceberg (less-so Delta Lake, but similar) is how difficult it is to just try out on a laptop. Delta Lake has vanilla Python implementations, but those are fragemented and buggy IME. Iceberg has just never worked locally, you need a JVM cluster and a ton of setup. I went down a similar road of trying to use sqlite/postgres+duckdb+parquet files in blob storage, but it was a lot of work.
It seems like this will just work out of the box, and just scale up to very reasonable data sizes. And the work from the DuckDB folks is typically excellent. It's clear they understand this space. Excited to try it out!
frisbm
Have you tried out PyIceberg yet? It's a pure Python implementation and it works pretty well. It supports a SQL Catalog as well as an In-Memory Catalog via a baked in SQLite SQL Catalog.
dkdcio
this looks much simpler than the last time I tried it!
akshayka
It's indeed very easy to try locally! For example in a marimo notebook, just a few lines of code: https://www.youtube.com/watch?v=x6YtqvGcDBY
(Disclosure, I am a developer of marimo.)
mritchie712
Here's a step-by-step setup. It's using S3 and RDS, but I wouldn't be hard to swap in a local sqlite instead.
TheCondor
I was thinking of putting something together for this. Like a helm chart that works with k3s.
datapains has some good stuff to get trino running and you can get a hivemetastore running with some hacking. I dorked around with it and then got the iceberg connector working with trino and see how it all works. I load data in to a dumb hive with a trino table pointed at it and then insert from select ... in to iceberg.
If the duck guys have some simple to get running stuff, they could probably start to eat everyone else' lunch.
wodenokoto
Delta-io (based on deltalake-r) runs very very easily locally. Just pip install, and write and you get catalog and everything.
dkdcio
I tried using it but on more than one occasion hit showstopping bugs -- they're probably fixed by now though
georgewfraser
They make a really good criticism of Iceberg: if we have a database anyway, why are we bothering to store metadata in files?
I don’t think DuckLake itself will succeed in getting adopted beyond DuckDB, but I would not be surprised if over time the catalog just absorbs the metadata, and the original Iceberg format fades into history as a transitional form.
wodenokoto
I’m building a poor man’s datalake at work, basically putting parquet files in blob storage using deltalake-rs’ python bindings and duck db for querying.
However, I constantly run in to problems with concurrent writes. I have a cloud function triggered ever x minutes to pull data from API and that’s fine.
But if I need to run a backfill I risk that that process will run at the same time as the timer triggered function. Especially if I load my backfill queue with hundreds of runs that needs to be pulled and they start saturating the workers in the cloud function.
isoprophlex
Add a randomly chosen suffix to your filenames?
ed_elliott_asc
Take a lease on the json file before you attempt the write and queue writes that way
kermatt
And a manifesto: https://ducklake.select/manifesto/
jakozaur
Iceberg competitor, addressing some of its shortcomings, like blown-up metadata:
https://quesma.com/blog-detail/apache-iceberg-practical-limi...
Even Snowflake was using FoundationDB for metadata, whereas Iceberg attempts to use blob storage even for the metadata layer.
buremba
I had the same impression but I wouldn't call it competitor after watching their video: https://youtu.be/zeonmOO9jm4?t=4032
They support syncing to Iceberg by writing the manifest and metadata files on demand, and they already have read support for Iceberg. They just fixed Iceberg's core issues but it's not a direct competitor as you can use DuckLake along with Iceberg in a very nice and bidirectional way.
prpl
metadata bloat can be due to a few things, but it’s manageable.
* number of snapshots
* frequent large schema changes
* lots of small files/row level updates
* lots of stats
The last one IIRC used to be pretty bad especially with larger schemas.
Most engines have ways to help with this - compaction, snapshot exportation, etc… Though it can still be up to the user. S3 tables is supposed to do some of this for you.
If metadata is below 1-5MB it’s really not an issue. Your commit rate is effectively limited by the size of your metadata and the number of writers you have.
I’ve written scripts to fix 1GB+ metadata files in production. Usually it was pruning snapshots without deleting files (relying on bucket policy to later clean things up) or removing old schema versions.
mehulashah
We've come full circle. If you want to build a database, then you need to build it like a database. Thank you DuckDB folks!
nehalem
I wonder how this relates to Mother Duck (https://motherduck.com/)? They do „DuckDB-powered data warehousing“ but predate this substantially.
raihansaputra
i think a way to see it is MotherDuck is a service to just throw your data at at they will sort it (using duckdb underneath) and you can use DuckDB to iterface with your data. But if you want to be more "lakehouse" or maybe down the line there are more integrations with DuckLake ir you want data to be stored in a blob storage, you can use DuckLake with MotherDuck as the metadata store.
jtigani
For what it's worth, MotherDuck and DuckLake will play together very nicely. You will be able to have your MotherDuck data stored in DuckLake, improving scalability, concurrency, and consistency while also giving access to the underlying data to third-party tools. We've been working on this for the last couple of months, and will share more soon.
nojvek
Motherduck is hosting duckdb in cloud. DuckLake is a much more open system.
Ducklake you can build petabyte scale warehouse with multiple readers and writer instances, all transactional on your s3, on your ec2 instances.
Motherduck has limitations like only one writer instance. Read replicas can be 1m behind (not transactional).
Having different instances concurrently writing to different tables is not possible.
Ducklake gives proper separation of compute and storage with a transactional metadata layer.
teleforce
Just wondering does DuckLake utilizing Open Table Formats (OTFs) since I don't see it's mentioned anywhere in the website?
buremba
My understanding was that MotherDuck was focusing on providing the "multiplayer mode" for DuckDB. It's interesting to see DuckDB Labs supporting data lakes natively. I guess MotherDuck is potentially moving to the UI layer by providing the notebook interface for DuckDB.
peterboncz
Good point! Anticipating official announcements I can confirm that MotherDuck is indeed intending to both: host DuckLake catalogs, and facilitate querying DuckLakes using DuckDB via its cloud-based DuckDB service.
spenczar5
There is a lot to like here, but once metadata is in the novel Ducklake format, it is hard to picture how you can get good query parallelism, which you need for large datasets. Iceberg already is well supported by lots of heavy-duty query engines and that support is important once you have lots and lots and lots of data.
buremba
You don't need to store the metadata in DuckDB; it can live in your own PostgreSQL/MySQL, similar to Iceberg REST Catalog. They solve query parallelism by allowing you to perform computations on the edge, enabling horizontal scaling the compute layer.
They don't focus on solving the scalability problem in the metadata layer; you might need to scale your PostgreSQL independently as you have many DuckDB compute nodes running on the edge.
spenczar5
Even though it's in your own SQL DB, there's still some sort of layout for the metadata. That's the thing that trino/bigquery/whatever won't understand (yet?).
> They solve query parallelism by allowing you to perform computations on the edge, enabling horizontal scaling the compute layer.
Hmm, I don't understand this one. How do you horizontally scale a query that scans all data to do `select count(*), col from huge_table group by col`, for example? In a traditional map-reduce engine, that turns into parallel execution over chunks of data, which get later merged for the final result. In DuckDB, doesn't that necessarily get done by a single node which has to inspect every row all by itself?
buremba
You are right, you can only scale vertically for single query execution but given that cloud providers now have beefy machines available with TBs of memory and hundreds of CPUs, it’s not a blocker unless you are querying petabyte level raw data. Trino/Spark is still unbeatable in that way but in most cases, you partition the data and use predicate pushdown anyways.
You can scale horizontally if you have many concurrent queries pretty well, that’s what I was referring to.
nojvek
you're correct that duckdb doesn't do any multi-node map-reduce, however duckdb utilizes all available cores on a node quite effectively to parallelize scanning. And node sizes nowadays get upto 192 vCPUs.
A single node can scan through several gigabytes of data per second. When the column data is compressed through various algorithms, this means billions of rows / sec.
formalreconfirm
Someone correct me if I'm wrong but from my understanding, DuckDB will always be the query engine, thus I suppose you will have access to DuckDB query parallelism (single node but multithreaded with disk spilling etc) + statistics-based optimizations like file pruning, predicate pushdown etc offered by DuckLake. I think DuckLake is heavily coupled to DuckDB (Which is good for our use case). Again, this is my understanding, correct me if wrong.
memhole
From my perspective the issue is analytics support. You’ll need a step that turns it into something supported by BI tools. Obviously if something like Trino picks up the format it’s not an issue
mrbungie
DuckDB has support for ODBC/ADBC and even JDBC (so Trino should be easy, https://trino.io/docs/current/connector/duckdb.html).
spenczar5
I agree with everything you said. I just mean that a single node may be slow when processing those parquet files in a complex aggregation, bottlenecked on network IO or CPU or available memory.
If the thesis here is that most datasets are small, fair enough - but then why use a lake instead of big postgres, yknow?
formalreconfirm
That's the part I don't really get. In the Manifesto they are talking about scaling to hundreds of terabytes and thousands of compute nodes. But DuckDB compute nodes, even if they are very performant, at the end are single nodes, so even if your lakehouse contains TB of data, you will be limited to your biggest client capacity (I know DuckDB works well with data bigger than memory, but still, I suppose it can reach limits at some point). At the end I think DuckLake is aimed at lakehouses of "reasonable" size the same way DuckDB is intended for data of "reasonable" size.
anentropic
It seems to me that by publishing the spec other non-DuckDB implementations could be built?
It's currently only DuckDB specific because the initial implementation that supports this new catalog is a DuckDB extension
snthpy
This looks very cool!
One thing I noticed is that the `table_stats` and `column_stats` tables aren't snapshot versioned. What are these used for and isn't that going to hurt timetravel queries (`SELECT COUNT(*) FROM tbl WHERE snapshot_id=<old_id>` as a simple example)?
BewareTheYiga
I am a huge fan of what they are doing, particularly putting local compute front and center. However for “BigCorp”, it’s going to be an uphill battle. The incumbents are entrenched and many decision makers will make decisions based on non technical reasons (I.e did my sales exec get me to the F1 Grand Prix).
mrbungie
Nothing new to see here really, pretty sure the same narrative was seen about Hive/Spark against Teradata/Exadata back in the 201X.
Wouldn't be surprised if 5 years down the line we see DuckDBaaS execs offering fully paid trips to DuckDBForce or smth like that.
I have an personal pet peeve about Parquet that is solved, incompatibly, by basically every "data lake / lakehouse" layer on top, and I'd love to see it become compatible: ranged partitioning.
I have an application which ought to be a near-perfect match for Parquet. I have a source of timestamped data (basically a time series, except that the intervals might not be evenly spaced -- think log files). A row is a timestamp and a bunch of other columns, and all the columns have data types that Parquet handles just fine [0]. The data accumulates, and it's written out in batches, and the batches all have civilized sizes. The data is naturally partitioned on some partition column, and there is only one writer for each value of the partition column. So far, so good -- the operation of writing a batch is a single file creation or create call to any object store. The partition column maps to the de-facto sort-of-standard Hive partitioning scheme.
Except that the data is (obviously) also partitioned on the timestamp -- each batch covers a non-overlapping range of timestamps. And Hive partitioning can't represent this. So none of the otherwise excellent query tools can naturally import the data unless I engage in a gross hack:
I could also partition on a silly column like "date". This involves aligning batches to date boundaries and also makes queries uglier.
I could just write the files and import ".parquet". This kills performance and costs lots of money.
I could use Iceberg or Delta Lake or whatever for the sole benefit that their client tools can handle ranged partitions. Gee thanks. I don't actually need any of the other complexity.
It would IMO be really really nice if everyone could come up with a directory-name or filename scheme for ranged partitioning.
[0] My other peeve is that a Parquet row and an Arrow row and a Thrift message and a protobuf message, etc, are almost* but not quite the same thing. It would be awesome if there was a companion binary format for a single Parquet row or a stream of rows so that tools could cooperate more easily on producing the data that eventually gets written into Parquet files.