I currently don't see the similarity to SQLite, that is implied in the readme.
SQLite stores all its data (every table and all entries in tables) in one single file. JSONlite does store every JSON entry in one particular file, as much I saw. So, JSONlite uses the file system (one directory) as data storage -- in contrast to SQLite!
How about some due diligence first? I know, it's so easy to say "just create a pull request". To which the answer is: "I'm not going to use your software if you can't make effort to test it and document properly".
Show me why your project is worth some of my finite time and effort, and I might. This looks like a toy project to me, and while there's nothing wrong with a toy project, I'm not seeing why it's more fun for me to play with this toy than some other of the myriad from which I have to choose.
Good point on 2 and 3. Should just involve checking the previous commands return code before echoing the uuid. I'll get that patched up and ship a new version in a bit.
With 1, you are trusting the directory returned by uuid will be unique --it should be, but it might not be, in particular if it is based on time and I run in two threads.
With 4, I find once a directory has about 100,000 files in it, things get bad, from the simple (ls * won't work) to the nastier (git starts using huge amounts of space, you'll hit github's size limit even though all your files are quite small).
I would not let anyone close to the code I'm working on if he wanted to verify whether a properly generated UUID clashes with some previously generated UUID.
You could just as do stuff like "int x=1; assert(x == 1);" and so on. And it would be just as futile.
In this case, you are relying on what installed program or function is called when you invoke uuid in the shell. It seems reasonable to me to check if it is having sensibly, if the reliability of your program counts on it. I have no idea what dodgy uuid programs you might find installed.
Edit: You may want more than 2 levels of directories, or directories with 2 character names instead of 1, etc. Apache's mod_cache defaults to 2 levels of 2 character names, but the filenames are base 64, so more possibilities than hex.
Heads up, just fixed the bug where providing invalid json to set would write an empty file with a uuid filename to the data directory. Now providing invalid json to set errors with no file written to the data directory.
Just wanted to say, glad to see these are being fixed. My initial message was a little harsh -- although github makes it hard to tell if you are making a new library for fun, or something you really want people using for important customer data!
This is 200 lines of bash using the file system as a database. The term document store implies a database which this is not. How is this different than just writing files to a directory?
The point is the project seems a bit low-effort. The most complicated thing in that bash script is argument processing.
What is the added value of that project? does it come up with something new or interesting? a query language for data? an efficient data storage? no, it just store json files in directories. Sure the file system is already a database, but that project doesn't add anything to the file system.
* Serverless typically implies something accessible from more than the host you're on
* It's not zero configuration, there's at least 1 configurable parameter (albeit with a sensible default)
* I'm actually not even clear why this is a document store limited to JSON, other than you're piping it though a JSON python module.
Having done some things like this in the past, as you continue, you'll probably want to create subdirectories based on the first N characters of the UUID, but I'm not sure you wouldn't get considerably more value of just putting your json blobs into an AWS Dynamo table.
I thought it was "serverless"? what does "sharding" have to do with anything? I'm sorry but it seems like you throw around some vocabulary that doesn't describe your project properly, to make it look bigger than it really is. That's nice marketing but it does not feel like there is much effort put into it, when I looked at your bash script.
The idea is breakup uuid keys into
sub directories based on the first couple of letters to prevent file system performance issues. Seems like the right usage of the word from a dictionary perspective.
I agree that "serverless" now mainly means "Function as a Service." The goal here is convey that there is no server component of the database, similar to SQLite and unlike PostgreSQL.
What is the correct terminology to highlight this architectural distinction?
Of course there is no server, since we are talking about local. I agree with others that the title is misleading. I looked at the project for a while to realize that it's just a local command line interface to storing objects in sqlite.
While I am sure there are use cases for this, it is not correct to say this is "serverless".
By that standard you can call literally ANYTHING "serverless".
For starters, this coffee I'm drinking right now is "serverless". It does its job well, completely without a server!
your headline is good except the term "serverless".
please drop it.
The term means much more than "not hosting it on a server". As somebody pointed out, "embedded datastore" may be a better term.
I agree the term is confusing now that "serverless" has a cloud-related definition, but in the database world, it is traditionally one of the terms used as an antonym to "client/server database". See, for example: https://sqlite.org/serverless.html
Serverless means anything really, a cli app is now serverless. I didn't know I was writing serverless applications for 20 years then. Serverless was already a bad buzz word, it gets more and more meaningless by the hour.
off-topic : I'd like to see more books on database implementation for beginners. Of all the crap load of CS books that come out year after year this is a matter with very little literature, as most papers on the subject are research papers. Seems like an excellent topic to teach distributed programming.
No, serverless database is not a recent buzz word, but a term frequently used to describe DB that doesn't need a separate server running to which clients connect: compare MySQL, PostgreSQL against BDB, LMDB, SQLite.
Not every serverless database is embedded, but most of them - those that are also called "database library" are. Some examples of DBs that are both embedded and call themselves serverless:
"serverless" means so many things it really means nothing, that's like "cloud computing" and that's the definition of a buzzword. Sqlite is an embedded database and has nothing to do with your favorite buzzword.
Hint: If you really want to use your filesystem as a database (and don't mind the haters, there's a long and storied history of doing just this), make sure you break it up into many subdirs usually based on the first few characters of the uuid.
Example: ./jasondir/aa/bb/cc/aabbccdd
You won't like what happens when you put 100k files in one directory.
Even if it's running on a filesystem that deals with it well, standard utilities like 'ls' don't deal with it well, as the default behavior is to sort. You end up having to look up obscure options, like 'ls -U' to disable sorting.
You can also run into issues like "Argument list too long". ARG_MAX is larger on linux than it used to be, but it's pretty short on older kernels. I assume similar issues might exist on other operating systems.
I'm not sure that's so much of a problem, I just tried:
mkdir uuid; cd uuid
uuid -v 4 -n 1000000 |\
while read uuid; do
touch $uuid;
done
And ran out of inodes, but:
time ls uuid|wc -l
425621
real 0m1.796s
user 0m1.552s
sys 0m0.240s
Sure, it's not exactly stellar performance for a linear scan of ~400k keys - but it's not terrible (for various values and expectations of terrible).
This is in a hyper-v vm on a Surface 4 pro/i5.
The fact that the "uuid" program can quickly generate uuids make me wonder if maybe one approach would be to generate uuids to (a) fifo(s), and then let db thread/processes read uuids from the other end?
Depends on the system, and 400k is short of where things tend to go bad on mine. At 800k files I see it taking 7 seconds.
Even at 400k files, you see longer wait times if you've aliased ls to ls --color (pretty common), or use something like ls -F. Either runs stat() on every file.
Then, somewhere in the 1m+ range, it gets unusable.
For troubleshooting or statistical reasons. It's serving as the backing data store. So piping through grep, or to wc, etc. Specifically not sending the output to the screen.
As this is a standard Ubuntu install, ls is indeed aliased to "ls --color" - but afaik ls as standard detects pipes, and turns off color (so you don't get a lot of control characters if you do "ls --color | sort > file.txt". Unless you use --color=always if I recall correctly.
Just wanted to differentiate between "select * from documents" and "select count(*) from documents" being slow.
I really don't get why this made it to HN first page, especially when plenty of other show HN demonstrate way more efforts than writing a short bashscript yet never make it. I suspect it has to do with the number of buzzwords that were inserted into the title.
I think there is a lot of feedback, mostly constructive criticisms. I would encourage you to embrace a lot of what everyone is saying and not to take it personally. You've done far better than most developers: you authored something and shared it with the world. To this day I'm still too afraid to show HN or Reddit my work.
First I want to say this is really cool. Thanks for taking the time to make it and I'll definitely give a test drive. That said, it's not zero configuration and your title is kinda misleading.
Why not just use SQLite? Its JSON1 extension is pretty good, and then you get all the rest of SQLite's goodness as part of the package. https://sqlite.org/json1.html
65 comments
[ 4.1 ms ] story [ 145 ms ] threadSQLite stores all its data (every table and all entries in tables) in one single file. JSONlite does store every JSON entry in one particular file, as much I saw. So, JSONlite uses the file system (one directory) as data storage -- in contrast to SQLite!
Instead, there are two tiny tests, and I found a couple of issues in 30 seconds of looking at the code.
EDIT: I should prove my point.
1) Race condition on calculating uuid (obviously won't be too serious for a good uuid implementation)
2) No check that file is written successfully.
3) No check when json is invalid, just silently swallowed, or filesystem full.
4) Makes one file a json file, will scale terribly past a few thousand json files.
Not sure what you mean exactly by 1 and 4 though.
For instance, if you run this database on Linux, the RAM overhead for each document is 1K just for the file handle, see the comment at the bottom here http://lxr.free-electrons.com/source/fs/file_table.c
With 4, I find once a directory has about 100,000 files in it, things get bad, from the simple (ls * won't work) to the nastier (git starts using huge amounts of space, you'll hit github's size limit even though all your files are quite small).
Many systems are relying on the statistical improbability of 32-hexdigits uuid collision (name-dropping one - Disque).
You could just as do stuff like "int x=1; assert(x == 1);" and so on. And it would be just as futile.
The usual workaround to that is to create subdirectories based on the first few characters of the filename.
Since you're using bash, you could do something like:
Edit: You may want more than 2 levels of directories, or directories with 2 character names instead of 1, etc. Apache's mod_cache defaults to 2 levels of 2 character names, but the filenames are base 64, so more possibilities than hex.New version is 1.1.0 (https://github.com/nodesocket/jsonlite/releases/tag/1.1.0) if you care.
What are the use cases of this project?
[1]https://en.m.wikipedia.org/wiki/Database
What is the added value of that project? does it come up with something new or interesting? a query language for data? an efficient data storage? no, it just store json files in directories. Sure the file system is already a database, but that project doesn't add anything to the file system.
* Serverless typically implies something accessible from more than the host you're on
* It's not zero configuration, there's at least 1 configurable parameter (albeit with a sensible default)
* I'm actually not even clear why this is a document store limited to JSON, other than you're piping it though a JSON python module.
Having done some things like this in the past, as you continue, you'll probably want to create subdirectories based on the first N characters of the UUID, but I'm not sure you wouldn't get considerably more value of just putting your json blobs into an AWS Dynamo table.
The idea is breakup uuid keys into sub directories based on the first couple of letters to prevent file system performance issues. Seems like the right usage of the word from a dictionary perspective.
What is the correct terminology to highlight this architectural distinction?
While I am sure there are use cases for this, it is not correct to say this is "serverless".
By that standard you can call literally ANYTHING "serverless".
For starters, this coffee I'm drinking right now is "serverless". It does its job well, completely without a server!
By the way sqlite just got support for JSON :
https://www.sqlite.org/json1.html
off-topic : I'd like to see more books on database implementation for beginners. Of all the crap load of CS books that come out year after year this is a matter with very little literature, as most papers on the subject are research papers. Seems like an excellent topic to teach distributed programming.
- SQLite: https://sqlite.org/serverless.html
- UnQlite: https://www.unqlite.org/features.html#serverless
Serverless is used to highlight this specific property of them. The OP's database is NOT embedded, but it is serverless.
Huh, what does that mean?
id=$(uuidgen)
echo '{"hi":"mom"}' | jq . > id
cat id
Example: ./jasondir/aa/bb/cc/aabbccdd
You won't like what happens when you put 100k files in one directory.
You can also run into issues like "Argument list too long". ARG_MAX is larger on linux than it used to be, but it's pretty short on older kernels. I assume similar issues might exist on other operating systems.
This is in a hyper-v vm on a Surface 4 pro/i5.
The fact that the "uuid" program can quickly generate uuids make me wonder if maybe one approach would be to generate uuids to (a) fifo(s), and then let db thread/processes read uuids from the other end?
Even at 400k files, you see longer wait times if you've aliased ls to ls --color (pretty common), or use something like ls -F. Either runs stat() on every file.
Then, somewhere in the 1m+ range, it gets unusable.
As this is a standard Ubuntu install, ls is indeed aliased to "ls --color" - but afaik ls as standard detects pipes, and turns off color (so you don't get a lot of control characters if you do "ls --color | sort > file.txt". Unless you use --color=always if I recall correctly.
Just wanted to differentiate between "select * from documents" and "select count(*) from documents" being slow.
I would be really happy to learn what are the use-cases for this tool, since using it via bash is actually an overkill.
1: https://github.com/yeoman/configstore
I really don't get why this made it to HN first page, especially when plenty of other show HN demonstrate way more efforts than writing a short bashscript yet never make it. I suspect it has to do with the number of buzzwords that were inserted into the title.