Sudheer Vinukonda created TS-3768:
-------------------------------------
Summary: implement a tiered back-off in read-while-writer retries
Key: TS-3768
URL: https://issues.apache.org/jira/browse/TS-3768
Project: Traffic Server
Issue Type: Bug
Components: Cache
Reporter: Sudheer Vinukonda
There's some code in the current cache layer that attempts to do a tiered back
off in reattempting read-while-writer (either for acquiring writer lock or
waiting for the first fragment download etc), but, that code is buggy (see
below).
https://github.com/apache/trafficserver/blob/master/iocore/cache/P_CacheInternal.h#L100
{code}
#define VC_SCHED_WRITER_RETRY() \
do { \
ink_assert(!trigger); \
writer_lock_retry++; \
ink_hrtime _t = WRITER_RETRY_DELAY; \
if (writer_lock_retry > 2) \
_t = WRITER_RETRY_DELAY * 2; \
else if (writer_lock_retry > 5) \
_t = WRITER_RETRY_DELAY * 10; \
else if (writer_lock_retry > 10) \
_t = WRITER_RETRY_DELAY * 100; \
trigger = mutex->thread_holding->schedule_in_local(this, _t); \
return EVENT_CONT; \
} while (0)
{code}
Presently, the retry happens as follows:
{{WRITER_RETRY_DELAY, WRITER_RETRY_DELAY, 2*WRITER_RETRY_DELAY,
2*WRITER_RETRY_DELAY, 2*WRITER_RETRY_DELAY etc}}
I'm removing the buggy code with TS-3767 to avoid confusion, but, opening this
jira per [~zwoop]'s suggestion to implement a proper tiered back off in the
future.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
