Running a Serverless DuckDB on Google Cloud

TL;DR : it easy to setup and works relatively well, but there is a catch, watch out for Cloud Storage throughput. I shared a notebook here

in previous blog, I showed a POC how to run Queries from a Colab notebook against a delta lake table, but what if you want to run the same Query from other tools, or if you want to run a Query in a different region and avoid egress fees, turn out it is extremely easy to setup.

And here is an Overall View of the architecture. the most important decision is to make sure Cloud Storage and Cloud function are in the same region, in my case “us-central1”, you can call the function from anywhere.

Google Cloud Functions

As I said before the code is very simple, I spent some time googling to convert the results from a byte to json to dataframe, I think beside BigQuery, Google Cloud function is the easiest service to setup, just write your code and Google Cloud handle the rest , just for fun, I used a machine with 8 CPU and 32 GB of RAM.

import pyarrow.dataset as ds
import duckdb
import json
lineitem = ds.dataset("gs://xxxxx/lineitem",format="parquet", partitioning="hive")
con = duckdb.connect()
def Query(request):
    SQL = request.get_json().get('name')
    df = con.execute(SQL).df()
    return json.dumps(df.to_json(orient="records")), 200, {'Content-Type': 'application/json'}

How to Call web API

a lot of tool can send a web api call as long as it has the correct authentication, I started with a python script for a simple reason as it was the easiest to get the code from the internet.

Here is the interesting part, you write any arbitrary SQL code, Then you send an API call, in return you get a json with the results, a user don’t need to know anything about the cloud function, all he needs is the web address and write a correct SQL Query.

Performance considerations

This is where it get very interesting, I have no clue where is the bottleneck, but we can ask some Questions

Cold Start

• The First example was 334 ms, that’s impressive, but I was cheating, I showed the best case scenario, Google Cloud function or more precisely Cloud Run was already running so no cold start and DuckDB was running a local Query which did not require a call to Cloud storage
• Currently Cold Start for Cloud run Gen2 is around 10 second, notice it is still in preview.

Transfer from Google Cloud Storage

let’s try this simple Query, we get the result in 29 second

The Same Query using my Laptop, 600 ms, btw table lineitem contains 60 Million rows !!!

I don’t know why the massive difference, I presume network speed is limited, but when I look at the the bucket stats, it is actually very good, nearly 500 MB/S

I am no expert in network, but that number don’t seems right !!, when I check how much Data the cloud function is receiving then the whole discrepancy start making more sense, in average I am getting around 30 MB/s, the maximum was 50 MB/s , I have to say this is really slow !!!

File Pruning

Arrow dataset is smart enough to prune columns that are not used for partitions, in this Query, I made a filter on L_shipdate , notice the parquet file was sorted on that field, and as expected the performance is very good 1.7 second, DuckDB scan only the row groups that contains the date ‘1998-09-02’

Dataset Catalog

I am defining a very rudimentary catalog, the user can just call ” Show Tables”

You can even check the table schema

Cloud Storage throughput is the bottleneck.

having a speed of 30 MB/S make the whole setup just good for POC or doing a Query on a small dataset , I don’t know the reason why such a poor performance from Cloud Run, I suspect Apache arrow implementation is not optimized for GCP although it works very well in a local file system.

Another missing piece is the lack of cache, it would have being good if somehow DuckDB cache the data already Queried, but cache is very hard to implement specially if you want cache invalidation, and you risk reinventing a full Data warehouse. I genuinely hope it is a bug and cloud run can provide a better network performance.