tilpner / betree Goto Github PK

Unsure if it works for cdylibs, let's find out

This will allow runtime flexibility w.r.t. per-object compression type, as opposed to the current per-database compression settings.
It is particularly essential to the C API, which can't monomorphise a DMU with different parameters.

Currently uses bincode, unstable format, and slower/larger than speedy

Empty keys will panic and keys with a size of many KB should probably be disallowed too

Questions

why bother with g_module_unload?

• special-case betree

integrate logging?

• betree_logging_enable (env_logger)

namespaces

threads from julea side

=> one thread per client conn

• (document assumption)

how to expose object listing

• tbd

whether to account for concurrent object aliasing

• yes => eager metasync

jemalloc in julea

when to write out file metadata?

• every time anything changes?
• on file close/drop?
• on every change

when to update metadata

• currently many times per write for consistency on partial writes
• good enough to update mtime once at begin and end of write?
• yes

sparse objects

• error or zero-fill when reading hole?
• how to configure? parameter of read_at, per-object, or per-objectstore?
• implemented

how/whether to integrate betree into build process

• no, sysadmin problem

Limitations

only one monomorphisation exposable for C

error handling is opaque

• translation into error categories

configuration from C is clumsy

• JSON/TOML path for config

partial writes/deletes due to lack of transactions

• no recoverability

on-disk format evolvability

• ignore

the entire cache and readahead is semi-pointless without direct IO

• keep both caches

fio/julea-benchmark or more realistic?

• https://github.com/enzo-project/enzo-dev

julea-fuse needs kv backend