The Neptune Shutdown:
We Have a Solution for You.

Neptune just announced the shutdown of its services.

Let us tell you why this might actually be a good thing. We have a very good deal for you.

We will help you save lots of time and money.

Drop-in Replacement: 1 Line of Code

We have a plug-in replacement for you. Same APIs. Just change 1 line and you're done.

Start today.

import neptune
import neptune_scale

import minfx.neptune_v2 as neptune
import minfx.neptune_v3 as neptune_scale

Data Migration

We will automatically migrate the data for you.

You don't have to do anything. Use all the features you're used to.

Minfx is Your Ideal Solution if ...

You are using machine learning in:

• Leadearboard-style situation, trying to make the best model, and/or
• You are doing continuous production retraining.

We are particularly aiming to optimize UI to make these as fast and easy as possible.

Expertise

We are experts in low-level optimization, both on backend and frontend, and care very much about performance and great experience. We have been building backend for a Neptune alternative, for a few months prior to Neptune's shutdown. And now we have a push when we want to make it available to you soon.

Frequently Asked Questions:

What will we get extra compared to Neptune?

We have the same API and feature parity as Neptune. So you can use your existing code with Minfx.

Additionally, we will have extra features:

• We offer a low-level CPU/GPU communication optimization that can save you ~10% of GPU costs.
• LLM-enabled inspections. We bring next-gen interfaces. You will be able to ask for insights about your metrics, runs, etc. in natural language. We then compile an appropriate program that fetches the data and analyzes it for you. You can edit these programs yourself as well in your favorite editor.
• Detailed dataset tracking: data versions, histograms of features, samples of data, etc. You can simply peek into samples of the inputs and outputs of your model to check it is what you'd expect.
• Detailed tracking of GPUs performance via nVidia's DCGM.
• GPU cost analysis: we help you to find ways to save money. Pick the right instance and batch size to find the pareto frontier of performance and time.
• Detailed learning monitoring: (per-layer) gradient norms, optimizer state, histograms of losses over time, etc. If you have gradient spikes, we will help you find the minibatches that are causing them.
• Multi-GPU support: multiple loggers per device for a single run.
• All values: We track all float values (inf/nan/+0/-0/signalling nans). We show them in the UI appropriately.
• High precision: We use nanosecond precision for timestamps.
• Time series anomaly detection: detecting anomalies in your metrics, runs, etc. We will search for things you might not have noticed.
• History revision: When you have a regression (in performance or accuracy) we will help you find when it started happening.
• Spot instances: Our logging works and tracks spot instance preemptions out of the box.
• And much more. Is there something you'd like to see? Let us know.

When will be Minfx ready?

Our service will be ready in Jan 30th, giving you a full month buffer. Plus, our migration tools make switching trivial. We will help you succeed.

What about other providers?

Well, there's not really much out there that can ingest data at the scale and speed needed for ML experiments we know of:

WandB: In our direct experience, long time ago they used to be a good choice, but now just don't care anymore. They were acquired by Coreweave, primarily in a move to get more customers for their cloud services. They recently redesigned their UI and somehow made it worse. Their servers now often return internal 500 errors. One word: decay.

Neptune, but that's not a choice anymore.

Then there's bunch of open source stuff, that you have to self-host to get any reasonable performance and deal with all the headaches. Such as: MLFlow, AIM, LabML, etc. And of course deal with migrations, rewriting your code, etc., all in the span of 3 months.

Why not just do it internally ourselves? LLMs are good these days.

Yes, they are for starting greenfield projects. But the complexity of the spaghetti code and lack of team's understanding of the code will bite you later on. There are "fun" things to deal with, such as:

• Setting up new infrastructure
• Corrupted data or data loss
• Hitting cloud rate limits
• Monitoring, alerting and detecting failures
• Cascading failures: What good is your ML training if your tracking layer fails?
• New inter-team coordination overhead
• and so on...

When these happen, what will you do? Do you have the people who are domain experts who can then dive in and fix the mess? This decision will cost you best-case tens of thousands of dollars.

Is there an API I can use today?

Not yet, but we are working hard on that. You will be able to use an API mid-January.

Do you have SOC2 certification?

No, we don't have SOC2 certification yet.

This still feels too much like a vacuous sales pitch. Is there any proof?

We are early. But we are very experienced engineers, and have been working on a very similar project prior to Neptune's shutdown.

We have lots of backend code already written, and we are now mostly working on hooking up to full Neptune's APIs, automating migration tools, and UI.

Who Are We?

Our team includes experienced ML engineers, who have worked at Tower Research Capital (algo-trading), processing petabytes of data at scale. We have strong academic and research credentials (PhD in Machine Learning, DeepMind). Finally, the team includes a founder of DataChine (real-time data processing), with deep expertise in building high-performance data systems and a strong focus on low-level optimization.

Progress

We are hiring!

Even though we have enough engineering resources, we are happy to talk with talented people to join our team.

If you stumbled upon this page and you ❤️ what we're building, reach out to us. Find the right e-mail address 🧩 to send your application. In the subject line write "Crate: <your-crate-published-on-crates.io-here>". This is a strong requirement, without it we won't consider your application.