Meta without dets #548
Meta without dets #548
Conversation
mkuratczyk
force-pushed
the
meta-without-dets
branch
from
9746b5a to
6c5117f
Compare
mkuratczyk
force-pushed
the
meta-without-dets
branch
2 times, most recently
from
218a8a7 to
4fa272f
Compare
Until now, we persisted metadata for all servers in a given Ra system in a single DETS file, owned by ra_log_meta gen batch server. This was a source of bottlenecks, so we move each servers metadata to a simple binary file. If the file doesn't exist, we try to fetch metadata from ra_log_meta and immediately persist it, so that we don't need ra_log_meta going forward.
Now that meta_fd is a part of the state, repeated calls to ra_server:init() lead to different initial states, so the test is adjusted not to expect the same result from both calls.
In this test we end up in update_last_applied with an undefined meta_fd set in base_state, so let's mock it as a no-op.
mkuratczyk
force-pushed
the
meta-without-dets
branch
from
4fa272f to
186d9f2
Compare
src/ra_server_meta.erl
Outdated
| NodeAtomBin = atom_to_binary(NodeAtom, utf8), | ||
| NameSize = byte_size(NameAtomBin), | ||
| NodeSize = byte_size(NodeAtomBin), | ||
| <<NameSize:8/unsigned, NameAtomBin/binary, |
There was a problem hiding this comment.
just noting that although an atom can only contain max 255 (possibly unicode) characters, the binary representation might require more than 255 bytes
1> Atom = list_to_atom(lists:duplicate(255, 256)).
'ĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀĀ'
2> byte_size(atom_to_binary(Atom)).
510
term_to_binary stores the atom size in 1 or 2 bytes https://www.erlang.org/doc/apps/erts/erl_ext_dist.html#atom_utf8_ext
sorry if this was considered already, and storing the size in 1 byte is intentional
There was a problem hiding this comment.
This was not considered.
I thought it would still limit utf8 atoms to 255 bytes. This is slighlty awekward as we'd theoretically need 4096 bytes just for the two atoms which would then take us over the standard block size.
We could consider compressing but we'd still need to ensure we're below 4000 or so bytes.
There was a problem hiding this comment.
Sorry it's late and I clearly cant math. We'd just need 2052 bytes to encode the atoms so we're still well below the block size.
There was a problem hiding this comment.
Thanks a lot for pointing this out. I've modified the format and code to use 2 bytes for atom sizes. Technically, if both atoms were full of UTF-8, we could still fail, but what are the odds?! :) Based on our testing, node names can't be UTF-8 (and the would need to be full of them for this to be a problem)
While an atom cannot be longer than 255 characters, its binary representation can be longer, if UTF-8 characters are present.
Deprecation of ra_log_meta with its DETS-based metadata persistence.
Instead, each server saves its own state in a simple binary file.
Of course for upgradeability, if a file doesn't exist, we take values from
ra_log_meta.