You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
--import download size is unbounded and v1 entries are read whole into memory #383
--import has no bound on download size, as in RSKj. It can fill the disk
before the SHA check fails, so the integrity check comes too late to protect the
host.
Separately, the v1 import format reads a whole entry into memory. On a 7.6 GB
box shared with the node, that is its own hazard — a 2 GB cache has OOM-killed
work here before.
Why it matters on this host specifically
/ has ~21 GB free and sdb is the build volume. An unbounded download that fills
either takes out more than the import.
Matching RSKj's behaviour is a defensible default for consensus, but this is not
consensus — it is operational safety, and we are free to be stricter.
What would settle it
Take the expected size from the index, refuse to start if the filesystem cannot
hold it with headroom, and abort the download once it exceeds the declared size.
Stream v1 entries rather than buffering whole.
Activity