Increase in client aborts with H2

[mlibbey]

We are seeing an increase of client aborts in http2 over http1 on 7.1.3:
proxy.process.http.err_client_abort_count_stat
proxy.process.http.transaction_counts.errors.aborts

• The vast majority of the aborts happen when we've sent 65535 bytes -- eg the "pscl" log. The object length varies -- eg, there isn't an obvious pattern to the object's content length.
• Most of the aborts happen for RAM cache hits.
• they get logged with cfsc="INTR" and crc="ERR_CLIENT_ABORT"

[shinrich]

Hmm. We saw a spike in client aborts too bringing in the 7.1.2 changes, but our issue was due to mislogging chunked responses. That was addressed in 7.1.3 by PR #3083

We are tracking the differences by analyzing squid logs not looking at the stats.

[maskit]

We probably need to define what client abort is. If a client close a connection abruptly (e.g. TCP Fin flag), that is obviously a client-abort from a perspective from HTTP transaction (== HttpSM). However, H2 client can send GOAWAY/RST_STREAM frame with NO_ERROR/CANCEL error code. Currently they're logged as client-aborts, and they both are technically client-aborts but maybe we should log differently or log the cause to distinguish bad ones from good(harmless) ones.

Also, I think client-aborts will be increased to some extent if we turn on H2 because transactions are multiplexed. That means one abrupt connection close causes more transaction aborts than H1.

[masaori335]

FWIW, TS can log Error Code in crec and ctec field by #3545.

[bryancall]

Looks like this has been fixed. Please reopen if there is still an issue.