luajit VS mumba

> No offence, but have you written any compilers or interpreters?

I have, but nothing sophisticated.

> The points that you discuss [...] may be performance concerns for application developers [...] but they have very little to do with the optimisations you can make as a compiler/interpreter writer. [...] The only one that's somewhat relevant is 'global scope by default'

I didn't mean to imply that these where the three common traits that make both Javascript and Lua particularly hard to optimize, I just picked them as examples for how Javascript and Lua are closer to each other than most other dynamic languages.

But let's dig in a bit on your claim that things like all numbers being doubles or having a array cum map cum record type has very little to do with the optimizations you can make as a compiler/interpreter writer, because it sure seems to me that LuaJIT and V8 do a bunch of optimizations around these things. Both have dual number representations under the hood and will try to avoid representing numbers that remain in the domain of 32 bit integers as double values internally when that gives performance gains. The logic for figuring out if that's the case doesn't seem to be super-straightforward or target architecture independent from looking at the comments in <https://github.com/LuaDist/luajit/blob/master/src/lj_opt_nar...>.

LuaJIT furthermore uses NaN tagging (as do some JS engines, although not V8), which looks less attractive to me as a representation strategy if your numbers are not all/mostly notional doubles (as is indeed the case in newer version of Lua where 64bit integers are the dominant number type)

Also, as far as the super-flexible lua tables are concerned, I'm pretty sure LuaJIT goes through some amount of trouble to specialize various common use cases of tables, e.g. as arrays without holes, and surprise, so does V8 (https://v8.dev/blog/fast-properties#elements-or-array-indexe...). I don't think you'd find something equivalent in a high performance scheme implementation.

> but this doesn't touch the surface of the issues that make JS hard to optimise, such as the fact that your, say, memoisation of an object property or method may be broken by an `eval` call of an arbitrary runtime value somewhere else in the code (which, due to asynchronicity, could take place at more or less any time from the point of view of your given 'peephole').

Eval belongs to a core set of features that basically all popular dynamic languages share that presents headaches for high performance implementations. How is Javascript's eval particularly problematic in this regard, and specifically much more so than Lua's loadstring/load?

More generally what do you think makes (pre-ES6) javascript significantly harder to optimize than lua 5.0?

> I feel like this would enable a more sensible choice:

I'm working in something like this

https://github.com/mumba-org/mumba

The applications developers publish native, Swift applications, where Swift have the full access to the Webkit/Blink api the same way Javascript does and even more with the patterns from the Chromium renderer (for instance, access to lifetime events).

The offline first comes from the fact that every application runs first as a daemon which provide a gRPC api to the world of the services it offers, where the web-UI application process might also consume this same api from its daemon -process manager or from other applications.

>That's interesting. Are you aware of the absurd-sql project?

No, thanks for pointing that out, will take a look into it.

> Also, how do peers find each other?

The idea is to use the torrents as a common shareable resource where given the peers have the same interest in that torrent, lets say the torrents works as a "meta-database" with just enough immutable metadata info, giving the developers of that application a list of peers that will have the same RPC service you designed, with a interface that suits your purpose according to your application goals (a distributed Youtube for instance), and giving you know and designed that API yourself whatever you want from each node, lets say a piece of data, be it a file or a database key-value range, you can ask your API for it, combining the torrent peers and whatever distribution combo you need.

> And finally, do you intend to open source it?

I've just did

https://github.com/mumba-org/mumba

It's badly documented given i'm on the final touches before a proper launch, but the "documentation" of the storage layer is planned for today, giving how important it is for the whole thing.

The first thing i've tried to get right was the storage layer, and the capacity to use mutable sqlite databases over torrent (together with files which are simpler given they are meant to be immutable).

Given theres no proper doc yet, i can point out to the source at

https://github.com/mumba-org/mumba/tree/main/lib/storage

where:

https://github.com/mumba-org/mumba/tree/main/lib/storage/bac...

is a modified chrome cache storage layer (which is the real underlying disk storage) that abstract the files and databases storages that from the bit-torrent layer perspective are on the disk.

https://github.com/mumba-org/mumba/blob/main/lib/storage/sto...

is the front end

and the:

https://github.com/mumba-org/mumba/blob/main/lib/storage/tor...

is the main abstraction that may be a "fileset"(collection of files as in torrent) or a dataset/database, which in this case you can get the sqlite db handle from the torrent object and deal with it as normal database.

I've take care to enable a key-value store over the sqlite btree, so both form of databases are possible, a key-value and a normal SQL database.

Key-values are important for the distributed case where you may want the nodes to have partial data and abstract a SQL layer (or whatever) on top of the distributed nodes, which is a better solution for distributed storage.

For the database distribution, a 64k SQLite memory page maps to a torrent piece of the same size, which can be synchronized over other peers that knows what's the root of the merkle tree of the given database is (you can use the RPC layer to coordinate this or use the bit-torrent DHT updating your slot with the new merkle root)

BTW This is what i'm using to distribute the applications, the DHT which points to a "database torrent" which in turn is a index to other files and database torrents.

I also have a project to explore this alternative way of peers communication but i have a different answer to this, and i think its better if its a network of peers that expose API's

https://github.com/mumba-org/mumba

It's badly documented as i have just published to github, but i hope it gives a clue of how is supposed to work.

I'm on the final touches over this project, but the main concept is already working as is 90% of it, but i think exposing SQL is too raw, and maybe dont offer the whole picture, as for instance, what is important is not data, but sometimes is computation..

Or yet, suppose you need to access something in a third-party before giving an anwer, or if you want to do it in a distributed fashion without you api consumer even noticing it?

API's are a good answer to that, and in my opinion are superior interfaces, whatever the semantic web of the future will be, it will need this network of API peers to work as a floor to it.

For instance, you can design a Graph API on top of it. Exposing your data layer directly is bad engineering as there's a lot of problems you wont be able to solve, ans where leaving clients to talk to "you" over a well-defined API will.

To put it simply, in my point of view the direction the semantic-web is pointing to is cool, but the answer is not the right one, and this idea of exposing SQLite directly while is cooler, yet have the same flaws, or else something as GraphQL would have taken the world as its not much a different answer than the one presented here.