When the proxy times out talking to a backend server (say, because it
was under heavy load and having trouble servicing the request), we catch
the ChunkReadTimeout and try to get the rest from another server. The
client by and large doesn't care; there may be a brief pause in the
download while the proxy get the new connection, but all the bytes
arrive and in the right order:
GET from node1, serve bytes 0 through N, timeout
GET from node2, serve bytes N through end
When we calculate the range for the new request, we check to see if we
already have one from the previous request -- if one exists, we adjust
it based on the bytes sent to the client thus far. This works fine for
single failures, but if we need to go back *again* we double up the
offset and send the client incomplete, bad data:
GET from node1, serve bytes 0 through N, timeout
GET from node2, serve bytes N through M, timeout
GET from node3, serve bytes N + M through end
Leaving the client missing bytes M through N + M.
We should adjust the range based on the number of bytes pulled from the
*backend* rather than delivered to the *frontend*. This just requires
that we reset our book-keeping after adjusting the Range header.
Change-Id: Ie153d01479c4242c01f48bf0ada78c2f9b6c8ff0
Closes-Bug: 1717401
6b19ca7a7d