Hacker News Reader

Comment by emersion | Hacker News Reader

>The single SQL endpoint is well suited for a data marketplace. Data vendors currently ship data in CSV files or other ad-hoc formats. They have to maintain pages of instructions on ingesting this data. With Splitgraph, data consumers will be able to acquire and interact with data directly from their applications and clients.

I appreciate the effort to make it easier for users to access heterogeneous data sets, but I really hope data vendors keep shipping raw CSV files. I don't want a company to gate access to the data, merely offering a proxy. I want to be able to download the whole raw datasets from the vendor directly if I want to.

View on HN

Absolutely, having ability to download the actual data and keep it is always going to be important. We want to facilitate access to data and think it should be available from the source. But, there will inevitably be fragmentation, so it's valuable to have a service available to catalog and aggregate it and make it available over a single protocol.

For what it's worth, we run PostgREST [0] on top of the DDN, so you can get your query results in JSON and CSV files. For example:

    $ curl -sSH "Content-Type: text/csv" https://data.splitgraph.com/cityofchicago/covid19-daily-cases-deaths-and-hospitalizations-naz8-j4nc/latest/-/rest/covid19_daily_cases_deaths_and_hospitalizations | wc -l

    172

It's limited to 10k rows but we might have an ability in the future to "order" a CSV dump asynchronously and place in a destination (like an S3 bucket) of choice.

[0] https://postgrest.org/en/latest/

Did you mean to use the "Accept" header? "Content-Type" describes the format of the request body.

Yeah, sorry, got too excited! It will return JSON by default and CSV with Accept: text/csv:

    $ curl -sH "Accept: text/csv" https://data.splitgraph.com/cityofchicago/covid19-daily-cases-deaths-and-hospitalizations-naz8-j4nc/latest/-/rest/covid19_daily_cases_deaths_and_hospitalizations?select=lab_report_date,cases_total,deaths_total | head -n5
    lab_report_date,cases_total,deaths_total
    2020-03-01,0,0
    2020-03-02,0,0
    2020-03-03,0,0
    2020-03-04,0,0

Very nice!

Any chance you’re willing to enable CORS on these endpoints? (Access-Control-Allow-Origin: * response header.)

Hey! Big fan of your work.

For the REST API endpoints (e.g. [0] [1]), we do set the Access-Control-Allow-Origin header to *, but for GET requests we do not send it at all. I'm not sure if this is the same as setting it to star. We can set it explicitly though; we do want to enable this use case (eventually we'll have quotas).

Soon we're releasing a Web UI for writing SQL queries. Part of that will include an HTTP (or websocket maybe) endpoint that takes a SQL query and returns the rows, so you'll be able to execute SQL queries directly from the browser.

As an aside, we'd love to get this working with Observable! Let us know anything we can do to help. You can email me at miles@splitgraph.com or join the Discord and we can figure it out :)

[0] Docs: https://www.splitgraph.com/cambridgema-gov/covid19-cumulativ...

[1] Endpoint (powered by PostgREST): https://data.splitgraph.com:443/cambridgema-gov/covid19-cumu...

Not setting it is equivalent to a deny

Ok, so this should be working now, for all types of requests. Let us know if it's not.

If you're testing with the same endpoints, you may need to give the cache a few minutes to evict the old values missing the header.

It works, thanks! https://observablehq.com/@mbostock/hello-splitgraph

In my view our collective interest in CSV as a medium for data distribution has resulted in far too much information loss, and consequently, time wasted on input sanitization, validity checking, and unresolvable conversations about the intent of data values like ”1.12345E+11” and "".

CSV is just an example, data vendors can choose e.g. JSON if they want more a sane and well-defined format.

Json has many design flaws, e.g. support for large integers, floating point infinity and NaN, for instance.

It is not a design flaw to make a reasonable choice about what data types you support. Especially given the ones captured in JSON are overwhelmingly the most commonly used and necessary types.

In cases where that's not enough, you could roll your own types by putting the values in plain strings and it would still be strictly more expressive than CSV

Regarding "sane & well-defined":

http://seriot.ch/parsing_json.php

Having spent way too much time wrangling vendor-provided CSVs, I 100% agree. I'd love for there to be a common, well-understood format for typed tabular data that supports multiple tables and enforces foreign keys between them. Ideally with a concept of "patching" to enable incremental updates.

Probably the closest thing I'm aware of is handing around a sqlite file, but I'm a little uneasy using a format that's meant to be a database as a transfer format. Dolt looks promising here too. Are there other ways?

https://www.sqlite.org/appfileformat.html - it is OK to use sqlite that way! =)

The Splitgraph core code on GitHub [0], around which we've built the DDN, is all about managing "data images" which are basically snapshots of PostgreSQL schemata. You can build them with a format similar to Dockerfiles as well as do a "checkout" into a local instance of Splitgraph (which you can connect to with any PG client) -- this enables change tracking and delta compression too.

Behind the scenes, we store them as cstore_fdw [2] files which is a columnar storage format that helps with analytical queries.

[0] https://github.com/splitgraph/splitgraph/

[1] https://splitgraph.com/docs/concepts/images

[2] https://github.com/citusdata/cstore_fdw

At work, we use Parquet (https://parquet.apache.org/) for almost everything related to a dataframe. We don't really care about performance gains (although, it's nice to have), but we really like to have a schema.

Note, we use mostly Python, some R, and a various range of ML or Optimisation tools, depending on the project.

Does Parquet provide a way to define foreign keys between different tables in a single dataset?

Parquet files don't reference anything outside of the file, usually. A group of parquet files in a folder is usually considered a table, where the schema is union of the schema of the files.

Parquet is kind of a royal pain in the ass compared to CSV/JSON/plaintext mostly because it uses a ton of Thrift encodings, resulting in mostly terrible/broken implementations anywhere outside of the Java/JVM ecosystem. If you're running Apache <Whatever> then sure, it'll probably be fine, but I'd recommend avoiding it if you start having to go down the rabbit hole of implementing support for things in your language du jour.

The Rust and python impl are fine. But I get it, Parquet may not be perfect or optimal or whatever. It works as a simple, typed, columnar format.

We had to pick a single file format recommendation for sending 100GB+ tables on FTP servers or dropbox, scanning terabytes of useless stuff only to grap an key-value pair, and properly reading integer and UTF-8 columns. Turns out, Parquet is practical. Enough for users to start using it instead of CSV. It could be Avro, but it's just not as easy.

> But I get it, Parquet may not be perfect or optimal or whatever.

I actually think Parquet is pretty great in practice, I just have some issues with the sheer volume of abstractions necessary to implement it. I just wish it was anything other than Thrift.

I would probably choose Parquet over anything else, though.

For exhaustivity sake, apache arrow could also do the work but Parquet seems more appropriate here https://stackoverflow.com/questions/56472727/difference-betw...

Avro is a strong contender for this use case: https://avro.apache.org/docs/current/

https://medium.com/ssense-tech/csv-vs-parquet-vs-avro-choosi...

It's well-supported in Pandas and Kafka, has good schema support, and reasonably small compressed file sizes.

Seconding parquet and sqlite, but hdf5 and netcdf4 certainly deserve a mention, especially for multidimensional, scientific, and generally large datasets.

See "CSV on the Web": https://www.w3.org/TR/tabular-data-primer/

The thing that attracted me to SplitGraph from the very start is that they are proposing to make the PostgreSQL wire protocol and SQL dialect a general interface to remote data. The interface is not only well known but backed by permissively licensed, open source libraries. Plus there are hundreds of tools that already connect to PostgreSQL.

This idea makes such sense it's a little surprising nobody did it before.

run-away queries become a problem when direct sql is exposed to public.

Both JSON and XML are self-documenting, and most modern databases support directly importing and exporting them. Though the tools to accomplish this could be better, these formats are far better suited as a "medium for data distribution" than CSV files are.

That said, a simple compressed .sql file of INSERT statements can often go a long way.

The only reason CSV is widely used is because normal people think of data as spreadsheets, and asking them to fire up a database and shred JSON data into it is ridiculous when they just want to whip up a line graph or answer a simple question (e.g. "What was value X on a this particular date?").

As an interchange format for database data, CSV is pretty terrible.

There's no comprehensive single specification. The closest is an RFC which completely misses major pieces (eg handling of binary data, and more).

Using SQLite as the interchange format is better. :)

IIUC, JSON is available too...

That's a good point. Any idea if there's a well spec'd JSON format around, for database data? eg something that handles binary data, trinary logic (eg null values), and hopefully referential integrity

You can generate CSV if you need it. See psql --csv. [1] What's brilliant about this approach is that you can generate any format that's supported by the interface defined by the PostgreSQL SQL dialect and wire protocol. (Obvious caveats about timeouts, network bandwidth, etc. apply.)

[1] https://www.postgresql.org/docs/12/app-psql.html

Reply on news.ycombinator.com