Serving a Million Models： How Hugging Face Scaled with MongoDB Atlas ｜ MongoDB.local London 2026

Good evening, thank you for joining my session. My name is Arek Borucki, and I work as a machine learning platform. I am a database engineer at Hugging Face. I have a question for you. Who knows Hugging Face? Good. Who uses Hugging Face? Perfectly. Today I'd like to show you how Hugging Face scaled up in MongoDB Atlas, And when we finished our duty almost three million models For developers around the world. I will divide the architectural decisions, challenges in the teaching which we have learned during MongoDB Atlas scaling in support of one of the largest developing communities of artificial intelligence around the world. I also have a live demo for you. I'll show you what Hub Hugging Face is I will tell you about MongoDB usage. Hugging Face. I hope you enjoyed this session. Let's begin. Before we move on to the technical details, Let's talk about scale. Right now, Hugging Face has more than 13 million users The tenth number is developing very quickly. We serve over 2 million public models. Almost three million I'll show you this during the live demo. We have more than 500,000 dates, more than 50,000 organizations They use Hugging Face. I'm not talking only about hobbyists or scientists. More than 30% of the Fortune 500 Uses Hugging Face in their AI services. Simply to put it into perspective. Two years ago, we had 20,000 models. We now have almost three million. MongoDB is our primary database. Almost all users on Hugging Face They use MongoDB. At this moment, we are using a set of replicas, And we will be operating our shops and using retail clusters, to be prepared for more users, More data, more development, and greater development. I want to show you a short video public models on Hugging Face Hub. The past two years have been amazing for us I kindly ask you to remember, that this slide stopped at 2025. Now we have 2026, What does it mean that we offer many more models? OK, let's now talk about high-level prediction. When a user interacts with the Hugging Face Hub, His proposal flows into Hub Frontend. after the Hub API and then to MongoDB Atlas. MongoDB Atlas is a source of real data our methods. This is an important point here, which sometimes surprises people. We do not maintain models in MongoDB. We keep everything about the models. What does this mean in practice? We use methods in MongoDB, user information, information repository And so on. Physical models are directly in S3 In MongoDB, only the point is stored for the model. in MongoDB documents Moreover, we index our model cards, Data sheets and certain application films with Python And we are setting it up in Atlas Search. This is what enhances our effectiveness. full-text search on Hugging Face Hub Now we will proceed to my laptop, But I want to show you Hugging Face Hub in practice. Now I'll do a live demo An introduction to Hugging Face Hub. We now have four pillars on Hugging Face, models, data, space, Bouquets, which are a new feature. I'll tell you about that right away. If someone wants to learn more, We have very good documentation. which describe everything in detail. You can also use Hugging Face Chat. to ask questions about this documentation. Now let's return to the models. One can use... Models can be searched for using many attributes. For example, I want to generate text I'll see how fast it is. In reality, I'm now searching in MongoDB Atlas, because the method is in MongoDB Atlas. For example, if I want to use fewer parameters, a few million less, I can still see the appropriate models. For example, I want to know, Which model uses transformers I almost have the probable results. Full-text search can also be used. to look for some information via Atlas Search. For example, audio text. I can go back to Hugging Face Chat and grab this model fairly quickly. So, this searching for the full text uses MongoDB Atlas for the purpose of. For example, if I want look for something about MongoDB in embedding Now I've returned to the MongoDB organization Hugging Face Chat. May I see the model card? I always have a description here. So, this model is good for example, for RUG services. I can fill out this form. I can use manufacturers Hugging Face Interface like a larger cloud producer. I can also use this model directly through Hugging Face transformers. So here I have a simple copy-paste to my code I will be able to use this model without much trouble. Let's go back to Hub. We also have data. So this place. Actually, I forgot to show you guys, We have almost 3 million models. Ten number changes very quickly. I am sure that in a few weeks It will be 3 million and something. Dane We have almost a million now. Again, I can use it very quickly Search with MongoDB I need to filter the models I require. If I go to Full Text Search again I will embed it in MongoDB. I will check the data. I have ready-to-use embeddings here. I am using these data from science So, for example, when I learned Atlas Vector Search I used Embedding in MongoDB from Hugging Face. Just because I didn't want to Generate embeddings yourself. The easiest part It was simply Get data from Hugging Face and send them to MongoDB. We also have a place. So this is the place for AI applications. Anyone can publish a place in this place. I would like to show you Something new. Our machine learning Inter. Actually, I have set up a MongoDB cluster. for this chatbot two weeks ago what does MongoDB mean It does not confirm only Hugging Face Hub. It confirms it. many other used cases. I'm here, for example May I ask, Who are you? That means this chatbot It will help me navigate. via Hugging Face search model. I think that too You can use this chatbot for connecting your own models. The last pillar Hugging Face Hub's surface Just Bucket. This is a new function. Maybe two months. And imagine, In which database Bucket is getting dressed. Of course, MongoDB Atlas, Now a cluster. The idea is that, AI agents and AI teams They can use these buckets, because they are higher from classic S3 things. They use Hugging Face's CDN in duplication technologies. Here I have an example, How to use these buckets. Maybe we should switch to the console. I can also interact with the Hub. via Hugging Face CLI I am already noticed. For example, I want to create a Bucket. I can use the CLI Hugging Face Bucket Create Let's put this into MongoDB.local. The bucket has been created. Let's try... Yes, it's here. If I go in Hugging Face's accumulator I should See that Bucket is there I can create my models in this Bucket I can use my agent to interact with this Bucket, because the agent needs it some data layer of interaction. Now let's move on to the slides. Now let's talk about infrastructure, Which one powers all of this. At this moment we are using classic replicas of MongoDB-7. W Daily Basic mamy ok. 40 kbps That number is also changing. Recently I saw disks up to 140 kbps, because our impact just now increasing And of course we must be ready for it. We use seven members of the replicants. Three of them They are electric. what it means that each of them can become the main one Three of them They are read-only discs. so we can isolate some heavy plates z electric plates. To do that, we'll just put it on connection keyboard near the key Which story? for reading only, then Your record will be balanced. for reading only We also use one analytical disc for some ad-hoc keyboards some analytical discs And so on. So we're trying no to mix our analytical traffic z electricity production tiles We are also trying to use preferences for reading, reading preferences private. So let's go on private discs, those who truly need them of strong consistency. Everything that is not needed strong does not have to use latest data is going to private. We also use in the near future. That means your disk may go to private and to private. It is a period after a period. So when you do tuning just You're checking your discs and you must make a decision, Is this disc Maybe it can go to private or maybe this record should go to private ones, private, or maybe this record I have to go to private. We are trying to use fast disks, SSD and also We have a few clusters, which they use local NVMe drives I am a big fan local NVMe because I know from experience, that the database is fast, if the database has Very fast and efficient disk, Of course. What is also useful in the Atlas, self-education The atlas is capable of distinguish the disk For example, if your disc received 80% 85% Atlas is capable automatically distinguish your association, and also Atlas is capable distinguish your cluster to the highest level. I saw it in production, sometimes it is useful especially when do we receive a few spikes, which occupy a few hours or perhaps a few days. So this is Today's architecture. The diagram shows seven replication sets. As I mentioned, we use basic of course for subtitles and sometimes for reading, Let's try anyway distinguish this traffic so many replication set members, How can we. Ten analytical This set is hidden set and if there's really something What is it some aggregation or something like that a very complex character I'm trying to use the hidden set for this type of employee just to make sure, that these characters are widespread from others traffic What do we do in secondary school? As I mentioned, We are trying keep our basics mainly for rights I if we have any comprehensive aggregations, They are going to the secondary school. and also to replication sets We are working with replication sets mainly in secondary schools, because we have a few sets, as thousands of sets, Some of them use replication sets and sometimes, when I am starting these sets, They are trying to connect. At the same time, to find one's place in replication sets, etc., so I decided Change these sets of the secondary schools They are separated. through I think, now five members, What makes our lives easier I ad hoc. Ad hoc Everything that is ad hoc, You're welcome. Sometimes people they are trying to investigate, What does not use replication sets, So I recommend them to you just join right away to hidden note I To run this kind of traffic there. Another important thing, which one we have in an ad hoc manner, data modeling Flexible schema It does not mean that there is no pattern. You need to spend some time on it and create an appropriate data model. For example, two weeks ago We started using MongoDB's computed patron so basically We're recalculating some data a bit. I Guidelines from the application use precomputed values instead of driving heavy traffic in our production database. We will also use sets. We use sets whenever we can. again One must remember, so as not to close too many things, then you will touch 16 megabytes limit and you shouldn't even approach this limit. So we should Insert into document Only information, which are truly needed, a potem MongoDB It works very quickly, because one query I accept everything. What do you need? to your application. Another important point to indices. Of course we use sets, but not only sets. Also multi-key indexes I We are trying to make sure those indexes They are really useful. What am I doing? Every 24 hours I am logging in. Z Atlas I am running log analysis For example, Keyhole I can see it, Which one is the query does not use indices in order list Which one is the query It doesn't use indexes at all. I can fix it. before We might run into some problems. So also I am observing use of indexes, because indices Of course they occupy space. and we shouldn't have more indexes, what we really need. Sometimes, as I have seen in other companies, and in the past as well, databases They had indexes, He doesn't use them, either. so I'm trying Extract such indexes. Actually for the first time I am using MongoDB Feature which allows me hide... I am using... You can hide the indexes, what does it mean that the query optimizer He will not see them. Then, if I don't see any bigger problems. I'll pull them out after two or three weeks. I Atlas Search It's a big deal for us. Now the full technical search is underway available In the Atlas region, and not only there. So I also noticed that MongoDB Community Edition supports the search for, what does it mean that it is separate MongoT process which consists of Apache Lucene and you can use effective full-text searching Also at the Premie. We don't use this At the Prize. Of course, we use We are happy with this atlas. Moreover, we have begun to use Vector Search which is not yet on production, I think we will use Vector Search also in production. Here we go. Sharding We're setting off right now Replica Set That's a simple setup. We have one basic, we have six secondary schools, Alright, let's go quickly. We are receiving a new wave of users Our databases are running. The oath is increased we decided to start using Sharding In the Atlas, as can be seen, Sharded Cluster has more components than Replica Set, So we have configuration servers, We have MongoDB databases. We have a balancer, etc. Fortunately in the Atlas It's easy to run Sharded Clusters. because Atlas is an automaton. I remember the time when I had to to rush to Prem. I take, That was possible. But it wasn't that easy, Nowadays At the Atlas. The only difficult part It's just selecting the Shard Key. This is the most important decision. the one you must take, When do you start? Use Sharded Clusters. Fortunately, now You can use shard key analyzers. That's what it's about I am working now. I am using MongoDB sampling What does it mean that I'm finishing? production waste for inclusion in MongoDB I do not use profiling. I'm afraid I'm worried, That will be too heavy for our clusters. That's why I use a sampler, which will not affect Performance is very important. Then I will use the Shard Key analyzers to analyze my ideas Shard Key I found this way Shard Key selections far easier than simply imagination. We also have something called Hugging Chat. This is another application, which one uses MongoDB In the Hugging phase. This is a chatbot. what Cloud yes ChatGPT could you choose your own open-source model who will be using this chatbot. MongoDB is the primary database. Where do we place all the methods? All settings and also conversations. So this is also atlas. Do you have a demo? I think we don't have time for this demo, I only have a minute, which was left. So I won't play this demo, But I want to tell you Mongolia. Mongolia is for MongoDB. We use this In the Hugging phase. You can check it on our GitHub. You can run this in your web browser. You can split the queries on Slack with your producers, Just by typing the link to Mongolia. We knew it very well useful and we use this in our online teams. So if you're interested, Just write down Mongolia And check this out. Thank you very much. It was pleasant for me. to be here. I hope my session was beautiful, and I hope so, that you are satisfied. Thank you.