I really like using fly.io for small projects. I like knowing cost up front though, so I wrote a CLI tool called flycost.
Like www.infracost.io but for fly.io. It reads your fly.toml file and gives a monthly cost estimate
I’d love feedback from anyone running on Fly.io
github.com/gordonmurray...
I built a simple to-do list app called www.notebare.com. All the data is stored in Lance format from @lancedb. Performance seems quick, Anyone can try it out, you can sign in with Github. Feedback more than welcome!
Cleaned up my old project: CDC from MariaDB → Flink → back into MariaDB using Docker Compose. Flink’s CDC connectors (powered by Debezium) keeps the target table up to date in real time. Code on GitHub: github.com/gordonmurray...
Apache Fluss is a storage engine:
• Fast key lookups (e.g. user status)
• Instant historic queries (via Paimon on S3/ORC)
• Works with Flink SQL for heavy lifting
• Tablets replicate & recover automatically
Made a small POC to try it out: github.com/gordonmurray...
I’ve refreshed a small hobby project from ~2 years ago: Apache Flink & Paimon.
It picked up a few early stars back then. I’ve cleaned it up so it’s runnable again.
I’m hoping to experiment with Apache Fluss on top of this. Code on Github github.com/gordonmurray...
I took some time this weekend to learn TLA+. It lets you test distributed systems by exploring all possible states. Meaning you can catch bugs before production and likely save costs. I built a small FastAPI + MinIO demo to try it out. Repo: github.com/gordonmurray...
I tried to build a minimal image vector store using Lance and Cloudflare R2. I Learned a lot, hit a hard limit fast.
"When Lance hits a wall at 70 images"
gordonmurray.com/data/2025/05...
Trying out Lance for storing vector data on R2. I’ve got Python saving the image and its CLIP embedding to a Lance dataset on R2. It’s append-only, columnar format kind of like Parquet, but optimized for updates, metadata, and search. Compared to Iceberg, feels like a better fit for vector mappings
I made a thing: upload an image, get back a vector embedding. That’s it.
Built with Python, FastAPI, and OpenAI’s CLIP model (ViT-B/32). Deployed on Fly.io.
quickvectors.com
Want to try it out? DM me for an API key.
Built a small DAM-style pipeline to handle image uploads, async processing, and ML enrichment. No DB, just new tools I’ve been meaning to try (fly.io, R2, Parquet). Wrote a post about what I’ve built so far and what’s next gordonmurray.com/2025/05/05/B...
Built a Python pipeline to upload images to R2 and generate captions + vectors using BLIP and CLIP from Hugging Face. Metadata stored as Parquet. Code on GitHub: github.com/gordonmurray...
Trying out Fly.io and Cloudflare R2 - long overdue on my part. Made a small Python app to receive an image, store it in R2 with Parquet metadata, and query via DuckDB+SQL.
Very basic, but next I want to try R2's data catalog / managed Apache Iceberg.
Code on github github.com/gordonmurray...
I created a small Terraform project that will set up Deepseek R1 for you on your own GPU EC2 instance, seems to be uncensored too. Code on Github at github.com/gordonmurray...
I took the steps from this great AWS Community guide community.aws/content/2sEu...
Updated my BentoML project to include Milvus for storing and searching vectorized images. More to learn about L2 Euclidean distance for better search results
github.com/gordonmurray...
A line diagram of making a Curl request to BentoML on ECS, protected by WAF
Just finished a small project on deploying BentoML on ECS using Terraform. Perfect for creating vectors from images with a CLIP model I’ve been learning. If you’re getting in to machine learning I'd recommend BentoML, hopefully this is a useful starting point!
github.com/gordonmurray...
Its surprising how many requests and countries try an API endpoint that has only just come online in the last few minutes, using a brand new domain name just registered too.
Tried out BentoML this evening to vectorize images, and it works really well! I expected to need a GPU, but it runs smoothly on a CPU. It even exposes Prometheus metrics out of the box. Code on GitHub: github.com/gordonmurray...
Ah this is great, I didn’t know this was available. The DDB approach was fine but this would be nicer
Re:Invent hasn't even officially started yet and we've already got 22 announcements. No rest on Re:Invent Sunday, I guess. To help you keep track, I will curate a number of custom feeds. The first one: All Re:Invent 2024 Releases
Check it out below 👇
The “show vehicle motion cues” option on iOS really works! For me at least. I can read away as a passenger and not feel queasy. Really cool 📖
If you’re thinking without writing, you only think you’re thinking.—Leslie Lamport
A screenshot of some JSON in a readme, using json5 so that comments can be added to the json snippet
The system works!
github.com/gordonmurray...
Oh nice. I didn’t know that. Thanks! I have some repos to go and update
Today is as good day to start learning Terraform
Hmm 🤔 Maybe I just created the AWS News Firehose on Bluesky. Maybe it's hosted at https://buff.ly/4g3kQsK 🦋 And maybe it's related to Bluesky's recent posts-as-comments demo... 🤷♀️ Who knows!
Very cool resource for learning about Apache Iceberg. I’d like something similar for Apache Hudi too
“Ultimate Directory of Apache Iceberg Resources” medium.com/data-enginee...
👍Ideally I’d set it up using Terraform though I assume it isn’t supported there yet.
I’ll set it up manually and see how it goes and write a post on it
I’ll give it a go too. I’m curious to compare it to using debezium and flink to do the same work
Ooh that’s kinda cool. CDC from MySQL directly to Iceberg on S3 using AWS Firehose.
I’m a big fan of debezium though this could be cool to try out. It’s might lend itself nicely to be managed in Terraform too.
aws.amazon.com/blogs/aws/re...
