With fleet state in LiteLLM, there is one artifact rather than two.
sudo -u postgres pg_dump litellm | gzip > "$dest/litellm-$stamp.sql.gz"
$dest is the open question. Three constraints:
- the destination must be captured off-host.
bpc_client_asf pulls via
rsync from a central BackupPC server, so the client has no local daemon and
coverage cannot be confirmed from the box. The share list lives on the
server — ask Infra which paths are pulled rather than assuming. A default
share of /etc and /home would miss /var/backups
- confirm the host is not in
bpc_client_asf::excludelist
- a file-level copy of a live Postgres data directory is not a valid
backup; the pg_dump is what makes it restorable
What is at stake, and it grew with "LiteLLM routes have no api_base": Postgres now holds routing as well
as keys, budgets and spend. Losing it means the proxy no longer knows what to
proxy, not merely who may call it. Virtual keys are stored hashed and shown
once at mint, so recovery also means re-minting every key and reconfiguring
every consumer.
With fleet state in LiteLLM there is one artifact rather than two, and nothing
outside the database to reconstruct from. The GPU boxes keep serving from their
local assignment, so a lost database does not stop inference — but nothing can
re-fetch, re-route or re-authenticate until it is restored.
With fleet state in LiteLLM, there is one artifact rather than two.
$destis the open question. Three constraints:bpc_client_asfpulls viarsync from a central BackupPC server, so the client has no local daemon and
coverage cannot be confirmed from the box. The share list lives on the
server — ask Infra which paths are pulled rather than assuming. A default
share of
/etcand/homewould miss/var/backupsbpc_client_asf::excludelistbackup; the
pg_dumpis what makes it restorableWhat is at stake, and it grew with "LiteLLM routes have no api_base": Postgres now holds routing as well
as keys, budgets and spend. Losing it means the proxy no longer knows what to
proxy, not merely who may call it. Virtual keys are stored hashed and shown
once at mint, so recovery also means re-minting every key and reconfiguring
every consumer.
With fleet state in LiteLLM there is one artifact rather than two, and nothing
outside the database to reconstruct from. The GPU boxes keep serving from their
local assignment, so a lost database does not stop inference — but nothing can
re-fetch, re-route or re-authenticate until it is restored.