Does the fact that you're still maintaining Kivaloo this many years later imply it's now being used by Tarsnap?
The Kivaloo page notes that it was designed for Tarsnap, but I'm curious as to the "why" - eg, an eventual-consistency/consensus building block, a simple local metadata service, server-side housekeeping...?
(I'm also curious what "{indexed ,}sorted files" means, and how UFS is significant.)
Do you mean with the OS file cache disabled?
Other questions:
1. What are, off top of your head, some design changes or code changes required that'd bring drastic performance improvements?
2. What are some key internals that you think differentiate Kivaloo from other embedded KV stores? I assume you must have gone through a lot of existing literature on the topic before building this. For example, LMDB, BDB, RocksDB, LevelDB, SQLite and the likes come to mind that can double-up as KV stores.
3. Does it store the database in flat files with a WAL in front? Is the file format of the database custom, or based on existing formats?
4. Does the database auto index the fields? Or, use any other such aids to speed up access to data?
Thanks.
No, I mean with a dataset which is too large to fit into the amount of RAM on the system.
1. What are, off top of your head, some design changes or code changes required that'd bring drastic performance improvements?
Nothing immediately comes to mind. Profiling may reveal some improvements, of course.
2. What are some key internals that you think differentiate Kivaloo from other embedded KV stores? I assume you must have gone through a lot of existing literature on the topic before building this. For example, LMDB, BDB, RocksDB, LevelDB, SQLite and the likes come to mind that can double-up as KV stores.
Well... kivaloo isn't an embedded KV store, so that would be a big differentiating factor. It's a network daemon.
3. Does it store the database in flat files with a WAL in front? Is the file format of the database custom, or based on existing formats?
The "on-disk" format is the pages of a append-only B+Tree, with the last page being the tree root.
I put "on-disk" in scare quotes because there are other backends, e.g. using Amazon DynamoDB to store pages.
4. Does the database auto index the fields? Or, use any other such aids to speed up access to data?
There are no fields. Key-value pairs, nothing more.
You might not get the same level of sophistication right away, of course, but that's just a matter of time and diligence.
A few questions:
- Why 255 bytes to 255 bytes? Does this have any performance consequences?
- Your laptop is SSD right?
- Is there any more documentation on how to use Kivaloo? (not just facts about it)
Yes, my laptop has a Intel 660p 512 GB NVMe disk.
No documentation per se, although I hope the library interfaces are reasonably understandable. I'd be happy to help though -- this code deserves to be used!
But if you're willing to make those adjustments and use 64 kB pages, I imagine it would work just fine. Please stay in touch!
This restriction could be relaxed, e.g. to require only that internal nodes are at least 2/3 full, in which case for the small key / large value case you could have pages which are barely larger than the largest value. I didn't bother doing that since it wasn't relevant to my usage.
For reference, on my laptop (Dell Latitude 7390 with an i7-8650U CPU):
* Bulk inserts run at ~600,000/second (up from 125,000).
* Bulk extracts run at ~660,000/second while in RAM (up from 30,000) and ~220,000/second from disk (up from 20,000).
* Bulk updates run at ~700,000/second in RAM dropping to ~300,000 from disk (up from 110,000 dropping to 60,000).
* Random reads run at ~800,000/second in RAM dropping to ~20,000 from disk (up from 220,000 dropping to 11,000).
* Random mixed run at ~400,000/second in RAM dropping to ~10,000 from disk (up from 30,000 dropping to 4,000).
* Hot-spot reads run at ~800,000/second in RAM dropping to ~500,000 from disk (up from 220,000 dropping to 60,000).
Note that "in RAM" means "the dataset fits into RAM" -- in all cases data is durably stored to disk.
There is almost certainly room for improvement; I haven't done extensive profiling yet.