We built a ChatGPT plugin for querying Splitgraph without writing SQL. Along the way, we discovered a better way to write LLM-powered applications.

How we built a ChatGPT plugin for Splitgraph


## Introduction

It seems like ChatGPT and similar large language models are appearing everywhere these days, from [web search](https://blogs.bing.com/search/march_2023/Confirmed-the-new-Bing-runs-on-OpenAI%E2%80%99s-GPT-4) to [coding](https://github.com/features/copilot), even in the [public sector](https://www.bloomberg.com/news/articles/2023-06-07/microsoft-offers-powerful-openai-technology-to-us-government-cloud-customers?leadSource=uverify%20wall).
One reason language models are so appealing is their ability to succinctly summarize and answer questions about large amounts of information such as <ValeIgnore>Wikipedia</ValeIgnore> articles.

That got us thinking: what if we could use a language model to answer questions about data in the [Splitgraph DDN](/docs/query/ddn)?
With over 40,000 public repositories, the DDN has a wealth of information available to anyone who can write SQL queries.
It turns out OpenAI's GPT is surprisingly good at writing SQL.
Does this mean Splitgraph can be used conversationally instead of writing queries?
The answer is yes, with a few important caveats.
We successfully built a proof-of-concept ([available on GitHub](https://github.com/splitgraph/splitgraph-llm-demo)), which can answer questions such as "How many public libraries operate in Chicago?" by prompting GPT to generate SQL queries which are then executed on Splitgraph.
Read on for the details!

## Introducing LangChain

In isolation, language models are only capable of generating or completing text.
In order to interact with an end user the way ChatGPT does or to use external data sources like <ValeIgnore>Bing's</ValeIgnore> web search index, OpenAI's GPT must be integrated with other systems.

The open source [LangChain](https://python.langchain.com/en/latest/) project aims to simplify the development of language-model powered applications.
Using LangChain, developers create pipelines (called "chains") which send text prompts to — and receive generated text from — language models.

We're not the first to try getting an AI to write SQL queries for us: LangChain includes an "SQL Chain" specifically designed for this purpose.
The [SQL Chain demo](https://python.langchain.com/en/latest/modules/chains/examples/sqlite.html) shows GPT is pretty good at figuring out which
tables to use in their query and composing a query to answer the original question in the prompt. The question is can it scale to meet Splitgraph's needs?

## The context is the bottleneck

**GPT** originally stands for [Generative Pre-trained Transformer](https://en.wikipedia.org/wiki/Generative_pre-trained_transformer).
Without diving into too much detail, transformers are a type of artificial neural network which process their entire input all at once.
Since their introduction in a [2017 paper by Google](https://arxiv.org/abs/1706.03762), transformers have taken the world by storm.

But processing all input "at once" would require arbitrarily large neural nets for arbitrarily large input.
For this reason, there is an upper bound on how large the input can be, called the maximum context length.

In order to process longer documents, they must first be segmented into text fragments which stay within this size limit.
The language model can operate on a single fragment at any time.
The LangChain SQL demo uses a database containing about a dozen tables. The entire database schema fits into the LLM's
prompt without reaching the maximum context length, meaning a single text fragment is enough!

With 40,000 repositories, each potentially containing several tables, there is no way to fit the DDN's contents into OpenAI's 4097 word (technically token) context.

Fortunately, having a huge collection of documents doesn't rule out using language models, after all Google or Microsoft integrated their entire web search engine indexes.

## Retrieval-Augmented Generation to the rescue!

[Retrieval-Augmented Generation](https://simonwillison.net/2023/Jan/13/semantic-search-answers/) is a technique for using LLMs with large collections
of documents. It uses a two-step process to answer a user's question:

1. First, it finds the documents most relevant to the original question.
2. Next, it generates an LLM prompt which includes the original question along with the selected documents (or their summary).

The trick is that the document search in step 1 doesn't require all the documents (or their summaries) to fit into the LLM's context.
Instead, search is performed by computing the vector similarity of the **embeddings** for the prompt and each document.

## Embeddings

An [embedding](https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/#the-concept-of-embeddings) is a vector of floating point numbers
generated by the transformer for a given input. The input could be a single token or a longer text as long as it's not longer than the maximum context length.

OpenAI's models use 1536 dimensional embeddings. In a sense, an embedding is like a hash function which computes
a list of floats for any text. Inputs whose computed embedding vectors are close together generally belong to text referring to similar concepts.

There are several vector databases which can be used to efficiently find the closest document embeddings to a given prompt's embedding.
Fortunately, we don't have to replace SQL, as the <ValeIgnore>[pgvector](https://github.com/pgvector/pgvector)</ValeIgnore> extension turns PostgreSQL into a vector database.

Needless to say, to compute the vector distance of document embeddings, we need to know the embeddings for each document -or rather each document fragment of maximum context size.
So prior to querying the set of all repositories, they must be indexed.

## Indexing repositories

The [indexing code](https://github.com/splitgraph/splitgraph-llm-demo/blob/main/index_repos.py) is fairly simple. Using the GraphQL API used by the Splitgraph frontend,
it's easy to get the data for each repository in a namespace. Since the README's for each repository use Markdown,
the entire description of the repository including the table schemas is generated in Markdown as well. LangChain has [document loaders](https://python.langchain.com/en/latest/modules/indexes/document_loaders.html) for lots of different text formats
including Markdown.

## Querying
OpenAI recommends using [cosine similarity](https://en.wikipedia.org/wiki/Cosine_similarity) when comparing embeddings, which `pgvector` supports with the `<=>` operator which is applicable to the vector type.
Once GPT generates an SQL query, it must be executed on the DDN.
Splitgraph implements the PostgreSQL network protocol, but some of PostgreSQL's introspection tables aren't available.
Normally, LangChain would read the database schema from these tables, but —lacking these tables— the [demo code](https://github.com/splitgraph/splitgraph-llm-demo/blob/main/langchain_demo/splitgraph_db.py#L93) uses the table schema returned by the GraphQL API instead.

## Results
We only tried querying repositories from one namespace, <ValeIgnore>[cityofchicago](https://www.splitgraph.com/cityofchicago)</ValeIgnore> during testing.
Occasionally, the SQL query returned by GPT can be invalid, the prompt can grow beyond the maximum context size for tables with lots of columns.
Still, overall, the generated queries are surprisingly good. Some examples:

```bash
> python3 query.py cityofchicago 'Which residential neighborhood of Chicago is the most expensive?'
SQL query generated by LLM:
  SELECT
    location_address,
    zip_code,
    units,
    property_name,
    property_type,
    community_area,
    address,
    community_area_number
  FROM
    "cityofchicago/affordable-rental-housing-developments-s6ha-ppgi"."affordable_rental_housing_developments"
  ORDER BY
    units ASC
  LIMIT
    5;
```

```bash
> python3 query.py cityofchicago 'How many public libraries operate in Chicago?'
SQL query generated by LLM:
  SELECT
    COUNT(*)
  FROM
    "cityofchicago/libraries-2020-circulation-by-location-v5qm-a67f"."libraries_2020_circulation_by_location"
```

One interesting thing about LLMs is that —depending on the temperature setting— their output can differ significantly
even if the input doesn't change.

## Conclusion

Large language models are capable of performing many tasks which were previously hard to automate —including the translation of natural language questions into valid SQL queries.
The LLM-generated SQL isn't always perfect, but it works surprisingly well surprisingly often.

Check out the [code on GitHub](https://github.com/splitgraph/splitgraph-llm-demo).
If you're excited about using Splitgraph with LLMs, [let us know](mailto:hello@splitgraph.com?subject=splitgraph%20llms)!


Follow along as we build a GPT-powered bot capable of answering natural language questions by finding relevant Splitgraph repositories and querying them via automatically generated SQL.

Building a GPT-powered agent to answer questions using data from Splitgraph


## Why Seafowl
Seafowl is a database designed to power data-driven Web apps like dashboards and interactive visualizations.
Its [cache-friendly HTTP API](seafowl-sql-cdn) is optimized for clients sending SQL queries directly from the browser.
As a single-binary database (built in Rust with <ValeIgnore>DataFusion</ValeIgnore> and [delta-rs](seafowl-delta-storage-layer)),
it's easily deployable to multi-region cloud functions that run "at the edge" and query columnar data in object storage.

Of course, before Seafowl can answer clients' SQL queries, it must have access to the data in question.
For Parquet or CSV data accessible over HTTP, Seafowl's [external tables](https://seafowl.io/docs/guides/csv-parquet-http-external) feature enables querying without any preceding import step.

For everything else, the data must be uploaded into Seafowl.
In this blog post, we'll demonstrate how to populate a Seafowl table using [Dagster](https://dagster.io), the hip new Pythonic framework that excels at creating declarative and reactive data pipelines.

## Why Dagster

Dagster solves the problem of unmaintainable data pipelines, by providing a framework for defining them with Python functions that you can run from the command line or a web UI. It doesn't require too much ceremony or "activation energy" to start using it, so even the simplest Python scripts or dbt jobs can start life as a Dagster [asset](https://docs.dagster.io/concepts/assets/software-defined-assets) and benefit from features like parallel processing of data on multiple nodes. You can add complexity if and when it becomes required.

If you want to run Dagster on your own, you can: it's open source. But Dagster Cloud is also available, so the same Python codebase can grow from a proof-of-concept on a developer's laptop to a production data pipeline powering applications gradually, without full rewrites.

## Inserting data into Seafowl from a Dagster pipeline

Here's how easy it is to append an "upload to Seafowl" step to the end of the Hacker News data pipeline from the [canonical "Hello Dagster" example](https://docs.dagster.io/getting-started/hello-dagster):

```python
@op
def export_to_seafowl(
    context, config: ExportToSeafowlConfig, data: pd.DataFrame
) -> None:
    context.log.info("exporting to seafowl")
    conn = SeafowlConnectionParams(
        url=config.url,
        secret=config.secret,
        database=None)
    destination = QualifiedTableName(schema="public", table=config.table)
    dataframe_to_seafowl(data, conn, destination)
```

Check out the [demo project repository](https://github.com/splitgraph/seafowl-dagster-demo) to follow along as we load the data used in the official [Hello Dagster](https://docs.dagster.io/getting-started/hello-dagster) example into Seafowl!

### 3-minute setup

#### Seafowl

If you don't already have a running Seafowl instance, head over to the [Seafowl Releases page](https://github.com/splitgraph/seafowl/releases/latest) and download the appropriate build for your platform.
Move the `seafowl` binary into the demo repo's folder prior to execution to use the provided `seafowl.toml` configuration.

#### Python and Dagster setup

You will also need a Python installation containing [Dagster](https://dagster.io/) and the [Python Seafowl client](https://github.com/splitgraph/seafowl/tree/main/examples/clients/python).

The following commands create such a virtualenv for you:

```bash
python3 -m venv venv
# Note: the following command is required before running anything installed in the virtualenv
. venv/bin/activate
pip install -e 'git+https://git@github.com/splitgraph/seafowl.git@main#egg=seafowl[pandas]&subdirectory=examples/clients/python'
pip install dagster dagit
```

### Starting Dagster
To start the Dagster UI (called "Dagit"), run the following:
```bash
dagster dev -f demo.py
```

Once the UI is running, the assets must be materialized before the data can be sent to Seafowl:

<video width="100%" height="240" controls>
  <source src="/static-v2/demo-videos/dagster-materialize.mp4" type="video/mp4" />
</video>


After the assets materialize, it's time to configure and run the `hn_stories_to_seafowl_pipeline` job:

<video width="100%" height="240" controls>
  <source src="/static-v2/demo-videos/dagster-run-job.mp4" type="video/mp4" />
</video>

### Querying Seafowl

The [Python Seafowl Client](https://github.com/splitgraph/seafowl/tree/v0.4.1/examples/clients/python) library can be used from the command line to verify the data has been successfully written to Seafowl by the Dagster job:

```bash
PASSWORD="iHBHflYfEaMpX4j7Du0z6vFNS6w2BMaX" ENDPOINT="http://127.0.0.1:8080/q" python -m seafowl "SELECT * FROM hn"
```

## Other ways to load data into Seafowl

As convenient as Dagster is, you might not need an ETL pipeline if your data is already in a Splitgraph repository.

You can export query results directly from Splitgraph to Seafowl. Just click the button with a database icon in the Splitgraph Console: ![Export to seafowl](/static-v2/seafowl/blog/seafowl-dagster/export-to-seafowl.png)

You can import CSV or Parquet files into Seafowl by [uploading them via HTTP](https://seafowl.io/docs/guides/uploading-csv-parquet).

In the above demo, the Dagster job invokes the [Python Seafowl Client](https://github.com/splitgraph/seafowl/tree/v0.4.1/examples/clients/python) to upload a Pandas DataFrame into Seafowl, which can be used directly from custom Python code.

For TypeScript users, the [madatdata library](https://github.com/splitgraph/madatdata) provides a polished API for querying and importing data into Seafowl.

Seafowl also includes a [Node.js client](https://github.com/splitgraph/seafowl/blob/v0.4.1/examples/clients/node/seafowl-client.js).

Finally, one can simply [`INSERT` data directly](https://seafowl.io/docs/guides/writing-sql-queries) into Seafowl.

The Python Seafowl client's `dataframe_to_seafowl()` function uses the Parquet upload endpoint if the `pyarrow` Python package is installed.
Otherwise, it issues `INSERT` queries for each row individually.

## Conclusion

Building interactive analytical applications like [Open Data Monitor](opendatamonitor) on Seafowl starts with uploading data. Dagster is a powerful tool for creating data pipelines which populate the Seafowl tables queried by users' browsers.


We're looking forward to seeing the incredible analytical applications you build with Seafowl! If you're excited about using Dagster with Seafowl or missing something, [let us know](mailto:hello@splitgraph.com?subject=seafowl%20dagster)!


Import the result of your data pipeline into Seafowl easily using Dagster!

Using Dagster with Seafowl


# Introducing SQLite file uploads

Splitgraph already had excellent support for [importing CSV files](csv-to-sql-upload-to-data-delivery-network-ddn), but now we can use the same interface to upload SQLite files as well. Try it out:

<EmbeddedUpload />

# Why use Splitgraph for SQLite data?

[SQLite](https://sqlite.org/index.html) is an incredibly versatile embedded database, used by apps on mobile devices, client code running in web browsers (via WebAssembly), and native applications.
The main difference compared to more traditional relational databases such as PostgreSQL is that an SQLite database is just a file, there's no need for a dedicated database server process.
This is an enormous advantage in resource-constrained environments and makes it possible to build cost-effective data-intensive applications using [Litestream](https://litestream.io/) for example.
At the same time, SQLite's database-as-a-file approach makes concurrent multi user access -and thus collaboration- difficult.

In addition to collaboration, Splitgraph offers countless features beyond what SQLite provides, for example:
- column-based data storage for efficient aggregation queries
- immutable data images
- built-in support for [`dbt` transformations](dbt)
- joining tables from other databases using [foreign data wrappers](foreign-data-wrappers)

# What about Datasette?

[Datasette](https://datasette.io/) is a versatile open source tool for data analysis and publishing built on SQLite.
It's a popular solution for making SQLite data available to the public, but doesn't have built-in support for controlling access to published data.
For those Datasette users who need [fine grained access control](data-sharing-settings-invites-private-repositories), Splitgraph offers a solution.

Similar to Datasette, Splitgraph also makes it possible to directly query database tables [from the browser](select-from-the-browser).
Unlike Datasette, authenticated Splitgraph-backed web apps can even write to database tables.

Datasette documentation includes two demo databases, [`global-power-plants.db`](https://global-power-plants.datasettes.com/global-power-plants.db
) and [`legislators.db`](https://congress-legislators.datasettes.com/legislators.db), both of which can be loaded into Splitgraph. You can also try the SQLite version of the [Chinook](https://github.com/lerocha/chinook-database/blob/master/ChinookDatabase/DataSources/Chinook_Sqlite.sqlite?raw=true) reference database. Try [uploading them here](/upload)!

# Key differences and limitations

SQLite [inherited](https://www.pgcon.org/2014/schedule/events/736.en.html) most of it's syntax from PostgreSQL, so the majority of SQLite queries should work on Splitgraph out of the box.
A major difference is that by default, SQLite allows any type of data to be written to any column, regardless of the column's declared type.
In fact, column type declarations are entirely optional in SQLite!

PostgreSQL is more strict when it comes to enforcing data types. Importing SQLite databases which store different types of data in a column than what was declared in the table schema may fail.
Tables containing columns without a declared type are ignored by the current implementation of SQLite import.

We'd love to hear from you! If you're excited about this new feature or if you're missing something from the current SQLite import functionality, [let us know](mailto:hello@splitgraph.com?subject=sqlite%20import)!


Importing SQLite data into Splitgraph is now as easy as drag-n-drop!

SQLite file uploads


## Measuring Urbanness

Sometimes a hard-to-measure property has significant impact on performance. For example, it's safe to assume sexier lingerie sells better,
but devising an algorithm to classify the "sexiness" of an article of clothing based on its design is far from trivial.

Another ill-defined quality impacting countless more measurable metrics is "urbanness", an individual's preference for densely or sparsely populated areas.
All sorts of trends, from [voting patterns](https://government.cornell.edu/news/exploring-widening-chasm-between-urban-and-rural-voters),
to life expectancy and religious affiliation all have marked differences in urban versus rural populations.

Statistically, it makes sense to classify someone as urban or rural based on their current residence, regardless of where they may have lived in the past.
On an individual level, it's not always so clear-cut. I currently live in a city with a population of over 1.5 million,
but I've also lived in a small town (less than 4000 residents). Today, a census would classify me as urban, but is that who I really am?

I decided to find out using photo location data. Smartphones usually encode the longitude and latitude in the EXIF data of photographs.
Over the years, I've collected thousands of photos (21808 to be exact), enough to give me a representative sample of where I prefer to spend my time.

## Classifying locations

The geometric center of a city is not always its most densely populated spot. What's a better proxy for "urbanness"?

If urban individuals are those who choose to spend their time in the city, areas offering many places to "hang out" are probably ideal locations for spotting urbanites.
Conversely, if one frequents such areas, chances are they can be labelled as urban. And what could possibly be a more urban place to hang out than the largest coffeehouse chain in the world, Starbucks?

Following in the footsteps of [Golden Arches Theory of Conflict Prevention](https://en.wikipedia.org/wiki/The_Lexus_and_the_Olive_Tree#Golden_Arches_theory),
let's define the equally tongue in cheek **Omnipresent [Siren](https://stories.starbucks.com/stories/2016/who-is-starbucks-siren/) Metric of Urbanness** (OSM) as the number of Starbucks locations within 2 kilometers of any given point on the globe.

Conveniently, there's a dataset of Starbucks locations [which can be downloaded from Kaggle](https://www.kaggle.com/datasets/starbucks/store-locations?resource=download), which I used to classify my photos by this metric.

## Loading photo coordinates

An easy way to extract EXIF metadata from images is to let Google Photos do it for you. For users of Google Photos, [Google Takeout](https://takeout.google.com/) offers a way to download photos along with their EXIF data extracted to separate JSON files.
Although not covered here, it's also possible to read location data (and much more) directly from images using [exiftool](https://exiftool.org/).

After downloading and extracting the Google Takeout archives, I ran the following to get a CSV file with the coordinates of each photo:

```bash
# enter Google Photos directory within Takeout directory
cd 'Takeout/Google Photos'
# extract image metadata from JSON files
find 'Photos from'* -name '*.json' -exec jq -c '{latitude: .geoData.latitude, longitude: .geoData.longitude, timestamp: .photoTakenTime.timestamp, filename: input_filename}' {} >> unfiltered_image_attributes.ndjson \;
# only keep image metadata with GPS data (non-smartphone cameras don't write location data)
grep -v '{"latitude":0,' unfiltered_image_attributes.ndjson  > filtered_image_attributes.ndjson
# write CSV header
echo "latitude,longitude,timestamp,filename" > photo_attributes.csv
# write CSV data
jq -r '([.latitude, .longitude, .timestamp, .filename]) | @csv' < filtered_image_attributes.ndjson >> photo_attributes.csv
```

## Loading data into Seafowl

Seafowl has great CSV support. To see it in action, [install Seafowl](https://seafowl.io/docs/guides/installing) if you haven't yet, then run:

```bash
curl -v -F "data=@starbucks_directory.csv" "http://localhost:8080/upload/public/starbucks_directory"
curl -v -F "data=@photo_attributes.csv" "http://localhost:8080/upload/public/photo_attributes"
```

Note that I changed the CSV header of the [Kaggle Starbucks dataset](https://www.kaggle.com/datasets/starbucks/store-locations?resource=download) for more convenient querying to:
```
brand,store_number,store_name,ownership_type,street_address,city,state,country,post_code,phone_number,timezone,longitude,latitude
```

## Calculating distances

Seafowl doesn't come with built-in functions for calculating the distance between two points on the planet,
but fortunately there is an excellent [Rust library](https://docs.rs/geoutils/latest/geoutils/) we can use.

To call it from Seafowl queries, we must create a user-defined function (UDF) and compile it to [WebAssembly](/blog/seafowl-wasm-udfs).
The code for the UDF is available [on GitHub](https://github.com/splitgraph/seafowl-udf-geo).

After running the [`create_udf.sh`](https://github.com/splitgraph/seafowl-udf-geo/blob/main/create_udf.sh) script, the previously unavailable `distance()` function may be used in queries
like so:

```sql
-- calculate the distance between Berlin and Paris in meters
-- distance(longitude_1, latitude_1, longitude_2, latitude_2) -> distance in meters
SELECT distance(52.5200, 13.4050, 48.8566, 2.3522)
```
We're ready to start classifying our photo collection by the Omnipresent Siren Metric!

## Calculating proximity: the naive approach

A simple way to calculate nearby Starbucks count for each photo is to find the distance to each Starbucks, then filter out distant Starbucks in the `WHERE` clause:

```sql
SELECT
    filename,
    COUNT(0) AS nearby_starbucks_count
FROM (
    SELECT
        filename,
        photo_lat,
        photo_lon,
        starbucks_lat,
        starbucks_lon,
        distance(photo_lat, photo_lon, starbucks_lat, starbucks_lon) as dist
    FROM (
        SELECT
            filename,
            latitude as photo_lat,
            longitude as photo_lon
        FROM photo_attributes
    ) p
    JOIN (
        SELECT
            latitude as starbucks_lat,
            longitude as starbucks_lon
        FROM starbucks_directory
    ) s ON true
) t
WHERE t.dist < 2000 -- distance is in meters, in this case a 2km radius
GROUP BY 1
ORDER BY 2
```

One of the great advantages of using an SQL database is that we can ask how a query will be executed with the `EXPLAIN SELECT ...` statement.
The logical plan for this query is:

```
Sort: nearby_starbucks_count ASC NULLS LAST
  Projection: t.filename, COUNT(Int64(0)) AS nearby_starbucks_count
    Aggregate: groupBy=[[t.filename]], aggr=[[COUNT(Int64(0))]]
      Projection: t.filename, alias=t
        Projection: p.filename, alias=t
          Filter: distance(p.photo_lat, p.photo_lon, s.starbucks_lat, s.starbucks_lon) < Float64(2000)
            CrossJoin:
              Projection: p.filename, p.photo_lat, p.photo_lon, alias=p
                Projection: photo_attributes.filename, photo_attributes.latitude AS photo_lat, photo_attributes.longitude AS photo_lon, alias=p
                  TableScan: photo_attributes projection=[filename, latitude, longitude]
              Projection: s.starbucks_lat, s.starbucks_lon, alias=s
                Projection: starbucks_directory.latitude AS starbucks_lat, starbucks_directory.longitude AS starbucks_lon, alias=s
                  TableScan: starbucks_directory projection=[latitude, longitude]
```

Notice the `CrossJoin` (also called Cartesian product)! It's the type of join between two tables where every row in table A is joined with every row in table B.
The resulting table will have a total row count of `rows in A * rows in B`.
The `distance()` function is pretty fast: a single call returns in 5-6 microseconds on my laptop. Even so, calculating all `25600 * 21808 = 558284800` distances took me about an hour.
Can we do better?

## Calculating proximity: using pre-filtered pairs

Calculating the precise distance between two points on a globe is more complex than on a plane.
Still, we can use a rough heuristic to filter out photo - Starbucks pairs which are obviously much farther away than 2 km.
At the equator, one degree latitude / longitude corresponds to 111km. I've never been so close to the poles that 2 degrees would be less than 2km apart, so any Starbucks
more than two degrees latitude or longitude away from the photo's location will surely be out of the 2km range.

### Pre-filtering pairs
First, I created a table of candidate pairs based on differences in latitude and longitude:

```sql
CREATE TABLE filtered_pairs_less_than_2_degrees_apart AS WITH
    all_pairings AS (
        SELECT
                filename,
                photo_lat,
                photo_lon,
                starbucks_lat,
                starbucks_lon,
                ABS(photo_lat - starbucks_lat) AS diff_lat,
                ABS(photo_lon - starbucks_lon) AS diff_lon
        FROM (
            SELECT
                filename,
                photo_lat,
                photo_lon,
                starbucks_lat,
                starbucks_lon
            FROM (
                SELECT
                    filename,
                    latitude as photo_lat,
                    longitude as photo_lon
                FROM photo_attributes
            ) p
            JOIN (
                SELECT
                    latitude as starbucks_lat,
                    longitude as starbucks_lon
                FROM starbucks_directory
            ) s ON true) r
    )
SELECT *
FROM all_pairings
-- A difference of 2 degrees or more latitude or longitude will definitely
-- result in more than 2 kilometers between the points in the range we're
-- working with.
WHERE diff_lat < 2.0 AND diff_lon < 2.0;
```

The `filtered_pairs_less_than_2_degrees_apart` table has 687911 rows, less than 1% of the rows resulting from the cross join in the previous query.

### Calculating the distance of candidate pairs
The following query filters the pairs further based on the precise distance:

```sql
CREATE TABLE filtered_pairs_less_than_2_km_apart AS SELECT * FROM (
    SELECT
        filename,
        photo_lat,
        photo_lon,
        starbucks_lat,
        starbucks_lon,
        distance(photo_lat, photo_lon, starbucks_lat, starbucks_lon) as dist
    FROM filtered_pairs_less_than_2_degrees_apart) t
-- Filter the list according to the actual distance in meters, so
-- < 2000 results in a 2km radius.
WHERE t.dist < 2000;
```

This second filter throws out over 94% of the rows in the `filtered_pairs_less_than_2_degrees_apart` table, leaving only
Starbucks locations within the 2km radius of each photograph.

The `filtered_pairs_less_than_2_km_apart` table contains 39715 rows, the same number we got with the original hour-long calculation, but the two
`CREATE TABLE` queries take a total of 15.1 seconds, less than 1% of the original runtime.

## My most urban photograph

There's a clear winner among the 21808 contestants for "most urban" photo, with an OSM of 68:

![Sheep in Shanghai](/static-v2/images/blog/20221209-seafowl-geo-distance-udf/IMG_6092.JPG)

Directing the [Starbucks Store Locator](https://www.starbucks.com/store-locator?map=31.232686,121.460572,15z) to the coordinates where the photo was captured helps explain why.
Considering Shanghai has the most Starbucks locations of any city on Earth, more than double that of Seoul at second place, it's not surprising that my "winning" photo was taken there.

## How urban am I

To find out, I needed the number of photographs for each OSM score:

```sql
-- query.sql
SELECT
    nearby_starbucks_count,
    COUNT(0) as num_photos
FROM (
    SELECT
        filename,
        COUNT(0) AS nearby_starbucks_count
    FROM filtered_pairs_less_than_2_km_apart
    GROUP BY 1
) t
GROUP BY 1
ORDER BY 1 ASC;
```

I used the following bash command to get the results of the query as CSV with the help of `jq`:

```bash
curl  -H "Content-Type: application/json" http://localhost:8080/q -d@- << END |
{"query": $(cat "query.sql" | jq -Rsa .)}
END
jq -r '[.nearby_starbucks_count, .num_photos] | @csv' > photos_by_osm.csv
```

Using the [Node.js Seafowl client](https://github.com/splitgraph/seafowl/blob/main/examples/clients/node/seafowl-client.js) is also an option:

```bash
seafowl/examples/clients/node/seafowl-client.js -f filtered_step3_simplified_osm_histogram.sql | tail -n +2 | jq -r '.[] | [.nearby_starbucks_count, .num_photos]' | @csv > photos_by_osm.csv'
```

With most of my photos having an OSM of 0, the answer seems to be that I'm not very urban, especially when compared to someone who feels at home in Shanghai:

![Number of photos by OSM](/static-v2/images/blog/20221209-seafowl-geo-distance-udf/number_of_photos_by_osm.svg)

## Urbanness over time

If "urbanness" is an innate quality unlikely to change much once our adult personality has developed, any reliable indicator should be pretty stable over time.
Looking at the number of photos I've taken with OSM ≥ 1 each month makes it seem like that is indeed the case.

![OSM ≥ 1 photo count by month](/static-v2/images/blog/20221209-seafowl-geo-distance-udf/osm_gte_1_photo_count_by_month.svg)

The exception is a trip to Shanghai in October of 2015 when I took the picture with the plastic sheep shown above.

## Conclusion

It was fun analyzing where I've taken photos over the last decade. Seafowl has been a convenient tool throughout the process:
- CSV import made it easy to work with data from [Kaggle](https://www.kaggle.com/datasets/starbucks/store-locations?resource=download) and Google Photos.
- We can easily extend Seafowl's builtin functions with [our own UDFs](https://github.com/splitgraph/seafowl-udf-geo).
- JSON output is convenient in most cases and can easily be converted to CSV using `jq` when necessary.

I look forward to all the cool stuff people build with this fast new database!


I used Seafowl to analyze how much of a city slicker I am—with geographic user-defined WASM functions and caffeine addiction!

Deciding if I'm urban with WebAssembly and Seafowl


## Do we really need User-Defined Functions?

User-defined functions (UDFs) are bits of application-specific logic which can be invoked within database queries, extending the builtin capabilities of the database.
You might be wondering why not keep application-specific logic in the _application_ layer? Are user-defined functions (UDFs) the relic of a past when databases were designed to be directly queried by end-users?

Far from it! Placing domain-specific logic in the database is often the best (sometimes only) way to implement functionality such as:
- Security policies such as PostgreSQL's [RLS feature](https://www.postgresql.org/docs/current/ddl-rowsecurity.html), which rely on UDFs to enforce application-specific security policies.
- Custom aggregation functions similar to `MIN`, `MAX` and `AVG` which must run within the database to avoid shipping the entire set of input rows to the client to calculate output.
- General, application-independent domain specific capabilities such as working with JSON fields, calculating geographical distances or full text search, extending the core DB's capabilities without having to modify the database itself.
- Finally, if raw and cleaned data live in the same database, it makes sense to use the database to transform that data with tools like [dbt](https://www.getdbt.com/) instead of passing it through a separate process -- and UDFs might be needed to implement some special functionality for that. [^1]

UDFs are an essential building block supported by the majority of databases one way or another, and in this regard Seafowl is no exception. What makes Seafowl special is its commitment to running "at the edge": as close to the user as possible, which means pretty much anywhere.
But if Seafowl is to run _anywhere_, UDFs referenced by Seafowl queries should be just as portable.
How can Seafowl offer UDF authors a platform which makes this possible?

## Platforms

Like any program, a UDF targets a specific platform. This platform should be hardware-independent so UDFs keep working if the database is moved to a different machine, but also efficient enough to process large datasets in acceptable time.

Mature databases often have a custom language for defining UDFs such as [PostgreSQL's PL/PGSQL](https://www.postgresql.org/docs/current/plpgsql.html) or [PL/SQL](https://www.oracle.com/database/technologies/appdev/plsql.html).
For simple functions just beyond the capabilities of an SQL statement, these languages offer an ideal solution.
For more complex functionality or code with lots of dependencies, many developers choose a general programming language, preferably one used in the application layer.

PostgreSQL and Redis both support authoring UDFs in Lua, for example. As a high-level language,
Lua abstracts away the lower layers of the stack such as the CPU's instruction set or the flavor of the C standard library.
Instead, the platform becomes the Lua virtual machine and standard library.
Lua is designed to be embedded in the host process, allowing scripts to efficiently invoke native functions exposed by the host.
It isn't the only scripting language embedded in databases. PostgreSQL for example supports many scripting languages including javascript and python.
While the source of the UDF may be portable, Lua, Javascript, Python, Ruby rely on native modules typically written in C or C++ for performance-intensive code such as crypto libraries.
These modules are co-located with the database and must be recompiled for each architecture supported by the database, limiting the portability of UDFs using these functions.

Custom logic written in compiled languages such as C++, Rust or Java is more difficult to invoke from a native database process, requiring
either running the UDF in a separate process or compiling against the C ABI of the database and executing it with full access to the database.
The security implications of the second option rule out this approach for UDFs.

Inter-process communication (IPC) with a separate JVM process provides a portable solution supported by [Oracle](https://docs.oracle.com/cd/B19306_01/java.102/b14187/chfive.htm) and [IBM DB2](https://www.ibm.com/docs/en/db2-for-zos/11?topic=java-stored-procedures-user-defined-functions),
but introduces a performance penalty for each UDF invocation along with additional corner cases arising from the need to synchronize a JVM process with the database.
For example, should updating the Java bytecode for a UDF impact running transactions?
If not, how many versions of the UDF must be kept in memory and how should the database keep track of which is the current version in each transaction?
When is it safe to remove classes upon which a previous version of the UDF depends?
If a transaction defining a UDF is rolled back, which classes' bytecode needs to be ejected from the JVM?
The list goes on and on, and it's mostly independent of Java, stemming instead from the marriage of a transactional database and another process which dynamically loads UDF code.

To summarize, there's hardly a silver bullet: each solution for providing UDFs has its pros and cons.
Worse, none of the solutions considered so far are truly portable in the sense that they can be used in the browser, on dedicated servers and desktops without serious limitations.

## Enter WebAssembly

Predating Seafowl and Splitgraph by decades, developers of large native applications faced a similar conundrum.
As new devices and platforms appeared, customers wanted to run their applications in previously unsupported environments.
Porting existing applications to new platforms was the traditional remedy, a slow and expensive process.

In the 90s, Java tried to address this problem, promising developers they could "write once, run everywhere".
As Java eventually found success in the very different niches of backend services and Android devices,
its early focus on "running everywhere" understandably changed to working best where it's used.

With the explosion of web apps, the browser appeared as yet another platform for applications to support, albeit a highly unusual one:
instead of a CPU instruction set, developers were to target a high level language, JavaScript.
Garbage collection came with the package, but integers were out (JavaScript has only floating point numbers).
Being so different from other platforms, porting existing applications to the web seemed more difficult than a complete rewrite in many cases.

Even green field projects struggled to find a stack sufficiently portable and efficient to ensure the application's long-term growth in desktop, mobile and browser environments.

As the performance of computers in general and JavaScript engines in particular improved, it became feasible to emulate a CPU in JavaScript.
A single C/C++ codebase could now be compiled to run natively on desktops or mobile devices, but also run in the browser with [emscripten](https://emscripten.org/), which targeted the javascript-powered CPU emulator.
It was hacky, but it worked and the world started to [take notice](https://kripken.github.io/mloc_emscripten_talk/#/).

With growing demand for efficiently running compiled code in the browser, Chrome, Firefox and Safari started to include new virtual machine called WebAssembly (WASM), very different from the high-level javascript VM:
- It wasn't optimized for a specific programming language like V8 or the JVM.
- It doesn't support sophisticated just-in-time compilation.
- It doesn't even have a garbage collector.

Instead, the WebAssembly specification provides a limited set of low-level instructions and linear memory model.
Thanks to the relative ease of implementation, today there are [many](https://wasmer.io/) [high](https://wasmedge.org/) [quality](https://github.com/wasm3/wasm3) [implementations](https://wasmtime.dev/), allowing code targeting WebAssembly to run
anywhere from embedded devices to [browsers](https://developer.mozilla.org/en-US/docs/WebAssembly), [CDNs](https://blog.cloudflare.com/announcing-wasi-on-workers/), and yes- in [databases](https://seafowl.io/docs/guides/custom-udf-wasm) as well.

## WebAssembly as a platform for UDFs

How does WebAssembly as a platform for implementing UDFs compare to scripting languages like Lua or more traditional compiled solutions like Java or C?
Since WebAssembly is designed to be used in very diverse environments, it cannot afford to make many assumptions:
- Memory is a sequence of bytes: there is no support for typed memory.
- As mentioned earlier, there is no garbage collector.
- WebAssembly code has no access to host process resources such as the filesystem, `stdout` or environment variables unless explicitly provided via functions that the WASM module can call at runtime.
- Catering to multi-tenant scenarios, hosts have find-grained control over how much computational power they grant a WASM function.
- The builtin types are limited to 32 and 64 bit floating point and integer numbers, bytes and booleans.

That last point presents a significant challenge when using WebAssembly for authoring UDFs. Many of the [types](https://seafowl.io/docs/reference/types) supported by Seafowl such as `TEXT` and `DECIMAL`
cannot be simply represented solely by scalar numeric values.

One option is to limit UDFs to consuming and returning types natively supported by WebAssembly (this is [ScyllaDB's approach](https://www.scylladb.com/2022/04/14/wasmtime/)).

For Seafowl's UDFs, we wanted users to be able to work with any datatype including strings, decimal numbers and dates.

In WebAssembly, memory is just a sequence of bytes and pointers are just 32 bit integers, similar to an array index.
Any object which can be serialized to a byte stream can be passed by the host to a WebAssembly function by first writing the serialized bytes to WASM memory,
then invoking the function with a pointer to the freshly written memory address as its argument.
Similarly, the WASM function can serialize it's output as a byte stream to which it returns a pointer.

Seafowl uses [MessagePack](https://msgpack.org/index.html) for serialization - think of it like a faster binary version of JSON.

## We're just getting started

Today, Seafowl supports creating UDFs in any programming language which compiles to WASM. Check out our [Rust example UDF repository](https://github.com/splitgraph/seafowl-udf-rust)! Any Seafowl type may be used for input values and return type.

One of the biggest advantages of using UDFs is aggregated records _within the database_. As of 2022 December, Seafowl does not yet support user-defined aggregation functions, but this support is planned. When it does arrive, you can be sure it will also use WebAssembly!

[^1]: There's an interesting discussion on this approach in this [Hacker News thread](https://news.ycombinator.com/item?id=33874842).


How can Seafowl users extend the database's builtin capabilities in a safe, portable and efficient way? The answer is WebAssembly, read on to learn how!

Extending Seafowl with WebAssembly


## Query Splitgraph from Public Observable Notebooks

[Observable](https://observablehq.com/@splitgraph) has native support for
connecting to [PostgreSQL](https://observablehq.com/@observablehq/databases)
from private notebooks and querying [SQLite](https://observablehq.com/@observablehq/sqlite)
files from public notebooks. Unfortunately, querying large amounts of data from
public notebooks has not been supported.

That's changed now thanks to
the [Splitgraph Observable Client](https://observablehq.com/@splitgraph/ddn),
which lets you query all the public data on the [Splitgraph Data Delivery Network (DDN)](/explore)
from any notebook.

## Import the client

To use the Splitgraph Observable Client in your own notebook, simply import it
from the `@splitgraph/ddn` notebook. Call the `splitgraph` function to send SQL
queries to the DDN over its HTTP API:

```javascript
import { splitgraph } from "@splitgraph/ddn"

all_cases = await splitgraph`
SELECT
  date,
  confirmed,
  recovered,
  dead,
  source
FROM "splitgraph/oxcovid19".epidemiology
WHERE countrycode = 'GBR' AND gid = '{"GBR"}' AND source = 'GBR_PHE'
ORDER BY date DESC`
```

## Example notebook

This [Planning a family vacation](https://observablehq.com/@splitgraph/planning-a-vacation-in-the-age-of-covid-19)
notebook showcases the Observable Splitgraph API:

<iframe width="100%" height="800" frameBorder="0"
  src="https://observablehq.com/embed/@splitgraph/planning-a-vacation-in-the-age-of-covid-19?cell=*">
</iframe>


Query public Splitgraph repositories from Observable notebooks, by importing the Splitgraph Observable Client to send SQL queries over HTTP to the Splitgraph Data Delivery Network (DDN).

Splitgraph has been acquired by EDB! Read the blog post.

By Peter Neumark

How we built a ChatGPT plugin for Splitgraph

Building a GPT-powered agent to answer questions using data from Splitgraph

Using Dagster with Seafowl

SQLite file uploads

Deciding if I'm urban with WebAssembly and Seafowl

Extending Seafowl with WebAssembly

Planning a vacation with Splitgraph and Observable

Product

Support

Company

Splitgraph

Splitgraph has been acquired by EDB! Read the blog post.

By Peter Neumark

How we built a ChatGPT plugin for Splitgraph

Building a GPT-powered agent to answer questions using data from Splitgraph

Using Dagster with Seafowl

SQLite file uploads

Deciding if I'm urban with WebAssembly and Seafowl

Extending Seafowl with WebAssembly

Planning a vacation with Splitgraph and Observable

Product

Support

Company

Community

Splitgraph