[HBASE-26469] correct HBase shell exit behavior to match code passed to exit

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 2.5.0, 3.0.0-alpha-2, 2.4.8
• Fix Version/s: 2.5.0, 3.0.0-alpha-3, 2.4.10
• Component/s: shell
• Hadoop Flags: Incompatible change

Release Note

<!-- markdown -->
User input handling has been refactored to make use of IRB sessions directly and the HBase shell attempts to ensure user provided calls to exit are able to convey failure and success.

Those scripting use of the HBase shell should be aware that the exit code may have changed:
    * a 0 code, or no code, passed to a call to exit from stdin in non-interactive mode will now exit cleanly. in prior versions this would have exited with an error and non-zero exit code. (note that in HBase 2.4.x this call will still result in a non-zero exit code)
    * for other combinations of passing in an initialization script or reading from stdin with using the non-interactive flag, the exit code being 0 or non-0 should now line up with releases prior to 2.4, which is a change in behavior compared to versions 2.4.0 - 2.4.9.

Please see the issue details for a table of expected exit codes.

• Assignee: Sean Busbey
• Reporter: Sean Busbey
• Created: 19/Nov/21 07:04
• Updated: 01/Sep/22 10:57
• Resolved: 15/Jan/22 20:51

Description

The HBase shell has changed behavior in a way that breaks being able to exit properly.

Two example scripts to act as stand ins for hbase shell scripts to "do something simple then exit":

tmp % echo "list\nexit" > clean_exit.rb
tmp % echo "list\nexit 1" > error_exit.rb

Giving these two scripts is possible:

• passed as a cli argument
• via redirected stdin

Additionally the shell invocation can be:

• in the default compatibility mode
• with the "non interactive" flag that gives an exit code that reflects runtime errors

I'll post logs of the details as attachments but here are some tables of the exit codes.

The clean_exit.rb invocations ought to exit with success, exit code 0.

	1.4.14	1.7.1	2.0.6	2.1.9	2.2.7	2.3.7	2.4.8	master
cli, default	0	0	0	0	0	0	1	1*
cli, -n	0	0	0	0	0	0	1	hang
stdin, default	0	0	0	0	0	0	0	0
stdin, -n	1	1	1	1	1	1	1*	1*

The error_exit.rb invocation should return a non-zero exit code, unless we're specifically trying to match a normal hbase shell session.

	1.4.14	1.7.1	2.0.6	2.1.9	2.2.7	2.3.7	2.4.8	master
cli, default	1	1	1	1	1	1	1*	1*
cli, -n	1	1	1	1	1	1	1*	hang
stdin, default	0	0	0	0	0	0	0	0
stdin, -n	1	1	1	1	1	1	1*	1*

In cases marked with * the error details are different.

The biggest concern are the new-to-2.4 non-zero exit code when we should have a success and the hanging.

The former looks like this:

ERROR NoMethodError: private method `exit' called for nil:NilClass

The change in error details for the error exit script also shows this same detail.

This behavior appears to be a side effect of HBASE-11686. As far as I can tell, the IRB handling of 'exit' calls fail because we implement our own handling of sessoins rather than rely on the intended session interface. We never set a current session, and IRB's exit implementation presumes there will be one.

Running in debug shows this in a stacktrace:

Took 0.4563 seconds
ERROR NoMethodError: private method `exit' called for nil:NilClass
NoMethodError: private method `exit' called for nil:NilClass
                     irb_exit at uri:classloader:/META-INF/jruby.home/lib/ruby/stdlib/irb/extend-command.rb:30
                     evaluate at stdin:2
                         eval at org/jruby/RubyKernel.java:1048
                     evaluate at uri:classloader:/META-INF/jruby.home/lib/ruby/stdlib/irb/workspace.rb:85
                      eval_io at uri:classloader:/shell.rb:327
     each_top_level_statement at uri:classloader:/META-INF/jruby.home/lib/ruby/stdlib/irb/ruby-lex.rb:246
                         loop at org/jruby/RubyKernel.java:1442
     each_top_level_statement at uri:classloader:/META-INF/jruby.home/lib/ruby/stdlib/irb/ruby-lex.rb:232
                        catch at org/jruby/RubyKernel.java:1189
     each_top_level_statement at uri:classloader:/META-INF/jruby.home/lib/ruby/stdlib/irb/ruby-lex.rb:231
                      eval_io at uri:classloader:/shell.rb:326
  classpath:/jar-bootstrap.rb at classpath:/jar-bootstrap.rb:194
            exception_handler at uri:classloader:/shell.rb:339
                       <main> at classpath:/jar-bootstrap.rb:194

And in our version of IRB (0.9.6) line 30 for extend-commands corresponds to this code:

    # Quits the current irb context
    #
    # +ret+ is the optional signal or message to send to Context#exit
    #
    # Same as <code>IRB.CurrentContext.exit</code>.
    def irb_exit(ret = 0)
      irb_context.exit(ret)
    end

Issue Links

• causes: HBASE-27086 graceful_stop cannot take previous balancer status by incompatibility of hbase shell prompt
• is caused by: HBASE-24772 Use GetoptLong or OptionParser in hbase-shell
• is caused by: HBASE-11686 Shell code should create a binding / irb workspace instead of polluting the root namespace
• relates to: HBASE-26741 Incorrect exception handling in shell
• relates to: HBASE-27116 Non-interactive missing for the hbase-shell
• links to: GitHub Pull Request #3901
• links to: GitHub Pull Request #4018