[
https://issues.apache.org/jira/browse/ARROW-12914?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Nic Crane updated ARROW-12914:
------------------------------
Summary: [C++] ArrowLog with FATAL level is not robust if running in the
service (was: C++ ArrowLog with FATAL level is not robust if running in the
service)
> [C++] ArrowLog with FATAL level is not robust if running in the service
> -----------------------------------------------------------------------
>
> Key: ARROW-12914
> URL: https://issues.apache.org/jira/browse/ARROW-12914
> Project: Apache Arrow
> Issue Type: Bug
> Components: C++
> Affects Versions: 1.0.0, 2.0.0, 3.0.0, 4.0.0, 5.0.0
> Reporter: Gang Wu
> Priority: Major
>
> We are a public cloud service provider and use Apache Arrow C++ as a
> dependency in the provisioned services. However, ArrowLog (CerrLog) w/ level
> ARROW_FATAL causes the service crash due to std::abort() call. The detail is
> as below.
>
> 1. The Arrow C++ implementation has defined marco ARROW_CHECK
> ([https://github.com/apache/arrow/blob/master/cpp/src/arrow/util/logging.h#L61])
> and uses it every where. If the condition in ARROW_CHECK does not meet, it
> creates an ArrowLog object with the level ARROW_FATAL and then logs something.
> {code:cpp}
> #define ARROW_CHECK(condition) \
> ARROW_PREDICT_TRUE(condition) \
> ? ARROW_IGNORE_EXPR(0) \
> : ::arrow::util::Voidify() & \
> ::arrow::util::ArrowLog(__FILE__, __LINE__, \
> ::arrow::util::ArrowLogLevel::ARROW_FATAL) \
> << " Check failed: " #condition " "
> {code}
>
> 2. The ArrowLog uses CerrLog
> ([https://github.com/apache/arrow/blob/master/cpp/src/arrow/util/logging.cc#L62])
> as the logging provider if GLog is not enabled.
> {code:cpp}
> #ifdef ARROW_USE_GLOG
> typedef google::LogMessage LoggingProvider;
> #else
> typedef CerrLog LoggingProvider;
> #endif
> {code}
>
> 3. The problem is in the destructor of CerrLog. It prints back trace and then
> dies with abort call. It results in crash of the service process and thus
> makes it unavailable.
> {code:cpp}
> virtual ~CerrLog() {
> if (has_logged_) {
> std::cerr << std::endl;
> }
> if (severity_ == ArrowLogLevel::ARROW_FATAL) {
> PrintBackTrace();
> std::abort();
> }
> }
> {code}
>
> I have traced back to https://issues.apache.org/jira/browse/ARROW-2138 and it
> seems that the behavior is expected. I'd suggest a more robust approach to
> get rid of the abort call. It should throw exception where the fatal error
> emerges and the logger's job is as simple as logging something. Thoughts?
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
