wat is GRASP?
i have started a negentropy peer implementation on orly now too, and it has a serve function that uses tmpfs, and blossom is default on, and respects whatever ACL system you enable. there is an aggregator tool that lets you trawl the nostrwebs for your events and writes them to jsonl. it has a GUI, which is still a bit of a WIP but already lets you log in, view and delete your own events. there is a zap-to-subscribe system for paid use case but it's untested as yet, same as blossom, though it's tested to the gills.
but wat is grasp?
Grasp is Blossom for git repositories.
You can host your repositories in multiple servers and their state is attested by Nostr events you publish that tells people where to fetch from and what is the latest head and branches, so the servers don't have to be trusted and it's easy to switch servers while keeping your repository identifiers immutable (they're essentially your pubkey + an arbitrary identifier).
Grasp servers can be self-hosted, paid, or ran for free by some community benefactor, and you can mix and use all these at the same time.
The existence of grasp servers also makes it easy for people to clone repositories, create branches and make them available to be merged by others, all without leaving the terminal (but of course there can also be all kinds of clients), which is the GitHub "pull request" UX people are familiar with.
Where is the spec for this?
📃.md
But I think it's still a work-in-progress. If you have some good ideas about how to best handle the pull request flow or other things let
@npub15qyd...yejr know.
Full Grasp support coming to BudaBit soon, the primitives are already implemented.
1. What we struggled with is the creation flow where we first post the repo announcement then the newly created git repo, but that must be done with an arbitrary sleep(X ms) wait time so that's not really ideal.
You guys have ideas to improve this flow?
2. I think it this was discussed before but want to get a fresh opinion of yall: Unified grasp api for file browsing and perhaps diffs and really data-heavy ops? Cloning repos via git smart http can still be a fallback but this would benefit performance a lot. Blossom has an api as well so I guess this would make sense, especially in a browser context.
@fiatjaf @npub15qyd...yejrFor 1, the 5s timeout used in ngit was more of a placeholder for a better approach. I think its best to pole the repo endpoint every 200ms until it reponds as if the repo as been created.
2. In the browser context we cannot rely on iso-morphic git as it doesn't support sparse clones and the UX of a shallow clone (with blobs) isn't good enough for large repos. eg. try browsing
GitWorkshop.dev
Decentralized github alternative over Nostr
for the first time.
I think we should explore directly requesting a pack from the http endpoint containing the blobs we need (for a specific file) instead of relying on iso-morphic git. We could create a javascript library that does just this.
If this doesn't work then we should add an API endpoint for files / listing directories, etc. My concern about the API is 1) it enables the use of a grasp server as a CDN for files in a repository 2) where would we stop in terms of the API, there isn't a clear boundary and we could end of creating all git commands / options as an API which makes it a more complex protocol and harder to implement.
Someone nearly attempted to add sparse clone in iso-moprhic git but there are a baked in assumptions in may parts of the codebase that blobs are present so it would require a larger change and it might be hard to get merged as a first time contirbutor.
Good idea thanks!
I dug into this a while back and hacked together a poc implementation. I wanted to store ojbects in blossom so I extended nip-34 to include content addressable objects.
you probably shouldn't do this but there's no bad ideas, right?
**Standard NIP-34 repository state (kind 30618):**
```json
["refs/heads/main", "c8c4e344a9c0b4008bb72eebb188be7d7b83dcb1"]
```
**extension for content-addressed storage:**
```json
["ref", "heads/main", "c8c4e344a9c0b4008bb72eebb188be7d7b83dcb1", "9adca865e86f755634dc0e594b74ba7d6ff5a66e4070c278e29ecda1bf7f4150"]
["obj", "589d02c42ef724523ceba0697315e36520332993", "abc123def456789012345678901234567890abcdef123456789012345678901234"]
["obj", "e63987bfc58e1197df38e5edb1d9930eb08d8589", "def456abc789012345678901234567890123abcdef456789012345678901234567"]
```
**Extensions:**
- **4th parameter in `ref` tags**: Maps Git commit SHA to content-addressed storage hash (Blossom SHA-256)
- **`obj` tags**: Map Git object SHAs (trees, blobs) to their content-addressed storage hashes
Interesting. You'd end up with a lot of objects with that approach and eventually it would be too big for the event size. I thought about doing it with storing packs in blossom. Here is my code to play with that idea.
GitWorkshop.dev
Decentralized github alternative over Nostr
I would have made it into a POC if rust-nostr had blossom support at the time. It does now. It turns out that having a git server is way more flexible so ngit.dev/grasp came to be. Let git be git and let nostr be nostr.
@arbadacarba are you worried that supporting so may ways of doing the same thing, adds unnecessary complexity, makes it harder to understand and harder to implement?
you could use Go and i already have a second draft blossom server written in go. i didn't write it. claude spun it up in about 3 hours and then another hour fixing it and i just haven't tested it yet. i know it works because it's just http and the tests pass, and i saw it accepting, and allowing me to delete a random blob several times. i just haven't used it. probably
will already serve you with blossom. imma make sure you both have permissions in case you want to try it
my take on this is look into techniques used in computer games. i remember when GTA3 came out, and its most epic achievement was loading free inter-map transit. still very few games use this but it's a graph theory algorithm.
this is the kind of thing you need to automatically, and quickly partition a map of related data. you need metrics of proximity and some kind of parameters for partitioning the map to fit the compute you need to do.
it's not hard. but it may take a while to wrap your head around it. but graphs at high node count are N! style compute cost. so it only takes like 3 or 4 deep and you are practically at infinity as far as even the most powerful computers can do in milliseconds.
yeah youd bloat the event with every object ref as the repo grows. not a great design but a fun poc.
i wrote my poc in go. the code is actually hosted on itself, as the poc is a relay/blossom/webui all in one binary server. i also wrote my own git-nostr-remote for the client side. it was a fun hack and generally works for the happy path. no planning to pursue it. i can share the code if you’re interested in it.
consider the option of using REQ and subscriptions. they are so fast if clients randomly put my relay wss://relay.orly.dev/ later than other relays, it has already received it via subscription using the follows spider. i will follow up this note with a screenie of it happening, if it happens
nope, it didn't do it, lets see if it works this time
humbug, imwald is sending it to my relay first.
anyway, it happens now and then, i see in the upload popup on the bottom right, that if my relay doesn't get sent the event first, it returns that it already has it.
The problem is git enables many feature like shallow and parse cloning, packing specific object and data, getting specific files and git logs, etc. These are all battle tested on solid git server implementations. This is all not possible trying to reinvent a simplified git server with blossom.
Nice! Would love to see it.
i'm definitely with you on that. i personally would even suggest to not even write one single bit of code that handles git. palm it off to Linus' excellent implementation and then see what it misses.
i always envisioned that nostr was just negotiation rendezvous easy inbound connection. i've already tried to work with an attempt to clone the core features of git, in pure Go, and it was endlessly problematic with even minimal lag between their work and what the git project has already progressed to.
git is just a unix shell protocol, based on stdio and unix filesystem. don't over think it. the protocol only needs to provide the correct references and paths.
also, it's just combinatorial stuff. you can't efficiently deal with multidimensional graphs, you want to stick with stuff that can be flattened into a 2d representation. branches and layers are basically exponentially more complex.
git is built on the directed acyclic graph geometry. there is a lot of shortcuts because you don't have to escape loops.
The reference implementation does exactly that. It uses nginx to pass request to the git-http-backend CGI scripts maintained by the git project.
I'm sure the midcurve meme apply here with 'just let git handle that'.
It turns out that the http wrapper (CGI scripts) is not that complicated. There are nearly 3 grasp implementations that do this bit internally. This enables shipping a single binary and handling the authorisation *before* the data is sent.
i dream of a day, where other people understand that most things are just pipes and shells and access control. if you ever read me on a regular basis you kknow the last point is my biggest gripe with most devs, and i already have experienced first hand how hard it is to explain the first two to most devs.
networks are just pipes with extra steps.
you mean, the retard and the jedi saying that, of course.
yeah, don't reinvent the wheel if you don't have to. servers that translate between protocol and git is quite trivial to implement. this is not realtime, low latency requirements here. even as much as 5 seconds to lay down a commit and propagate it is fine.
never prematurely optimize, and don't roll your own if you can just assemble someone else's stuff into a shape that meshes with your actual part of it.
speaking from experience. even since picking up LLMs to help me with this, there are many things that just are not practical to do of such scale in any sane amount of time.
it was a harsh burn discovering that i couldn't do git stuff in pure go. the go version just isn't nearly adequate. it seemed to be working, and then i got all this mess going on. what was the error? i forget, it was some protocol network shit iirc. annoying af.
thats' why
git.mleku.dev
ᴛʜᴇ ᴅᴇᴀᴛʜ ᴏꜰ ᴍʟᴇᴋᴜ
Gitea (Git with a cup of tea) is a painless self-hosted Git service written in Go
is just a plain gitea. i wanted to just host only my repos, and not have that stutter in the URL.
i spent probably weeks trying to get that working, and in the end, it was futile. linus seems to be turning into a javascript ninja with all his fucking breaking changes this last few years. fuck that guy.
just use his binary and interface to the git repo using it.
the bitch has got too complicated to build from scratch. oh, sure, we could do all kinds of things involving metadata on nostr and all that shit but you really should just pause, before you race off and do that, and go clone the git repo, and tell an LLM to explain it to you. 30 screenfuls of documentation later, you will be in agreement with me.
nah, just call git via your preferred language method for executing child processes. the end.
Not anylonger but it was certainly a fight. Made a PR on the optional changes back to gitnostr and a repo on snippets with which stuff got better to handle. And serious doku, not the unuseful kind but actually helpful 🐈⬛
If everyone contributes back and new ones still know where to start we' re good 🤗
Issues and PRs (kinds 9803/9804) are automatically published to nostr on handled status changes (merged, closed and reopened).
I fetch them from source if possible on import of the repo and try to aggregate those by their timestamps with the nostr kinds.
If source is lets say Github im not upstreaming the edits additionally there so far.
Anyway still needs polish in finding these kinds better and flows are surely not the endgame, but what i went with so far 🤓
I'm a bit confused. Are you are saying the extra complexity of having lots of different ways of doing the same thing is OK? And that you made a PR to a project that has only 1 commit posted 2 years ago.
Yes, thats what i am saying. Its about discoverability for me.
See nothing wrong in PRing changes so its up to date. Lost and left repos are a problem when ppl want to contribute.
The result is very reliable & fast so i dont really get the problem here neither with complexity and sharing snippets on that nor PRing progress. Ive contributed to much older code than 2 years before..
Id like to know what's too complex in this. You can just fork it and build upon or just face the complexity.
release (v0.17.0) has:
- `nak git` commands that allow cloning, setting up a new nip34/grasp repository, pushing, fetching and pulling (just call "nak git push", for example, instead of "git push")
- `nak req --only-missing` flag that takes a jsonl file with events and does negentropy with a target relay to only download the events that are not in that file (finally @Kieran -- this was ready 2 weeks ago, but I had to make a ton of git stuff before I was able to publish it)
- `nak serve --negentropy --blossom --grasp` new flags that make hosting these things of servers locally much easier for debugging
- you can finally use npubs/nprofiles/nevents/naddrs directly as arguments to `nak event`, `nak req` and others (they will be parsed and included in the event or filter as proper hex)
