Proposal to change the behavior of the timestamp place holder (%t) in log file paths

[Laurence Cable]

On 1/29/25 5:56 PM, David Holmes wrote:
> Adding back serviceability-dev
>
> David
>
> On 30/01/2025 11:55 am, David Holmes wrote:
>> Plus one to what Kevin says. The current specified behaviour seems to 
>> me to be what would be generally desired. If there is a desire for a 
>> different kind of timestamp to be used then a proposal to add that 
>> would be more appropriate I think. I seem to recall something fairly 
>> recent about expanding the notion of "timestamp" that can be used here.

+2 for Kevin and David's observations; changing the timestamp from JVM 
start time to create time, removes valuable and
otherwise unobtainable (correlation) metadata from the logfile names, 
file creation and modification time is available from the
underlying O.S filesystem if needed.

Rgds

- Larry

>>
>> David
>>
>> On 29/01/2025 7:24 pm, Kevin Walls wrote:
>>> Hi,
>>>
>>> Just checking, but are they sure that's what they want? 8-)
>>>
>>> This relates to files from unified logging, like java - 
>>> Xlog:gc*:file%t.out ...creates:  file2025-01-28_21-43-53.out and - 
>>> Xlog:help says, "If the filename contains %p, %t and/or %hn, they 
>>> will expand to the JVM's PID, startup timestamp and host name, 
>>> respectively."
>>>
>>> (Administratively, I think unified logging is under the runtime 
>>> group, cc’d.)
>>>
>>> Using the JVM start time, across all log files, identifies the set 
>>> of files that come from the same process.  They will generally sort 
>>> together when viewing a directory.  If a single file gets copied 
>>> around, it can still be traced back in its group.   When there are 
>>> multiple sets of logs in the same directory, the sets should still 
>>> sort together.   I see the filename purpose as to identify the log, 
>>> or set of logs.
>>>
>>> Using a new timestamp for each file, the filenames do not identify 
>>> the files as being part of the same run.  They may sort together, 
>>> but may not if another log series is in the same directory, and once 
>>> separated it's hard to regroup them.
>>>
>>> Using the pid as well will help, but if we see a lot of low-numbered 
>>> PIDs then this won't be unique. (With the current startup timestamp, 
>>> you will probably use %p for pid in the file as well, in case JVMs 
>>> start at the same moment.)
>>>
>>> Thanks
>>>
>>> Kevin
>>>
>>> *From:*serviceability-dev <serviceability-dev-retn at openjdk.org> *On 
>>> Behalf Of *Kemper, William
>>> *Sent:* Tuesday, January 28, 2025 7:54 PM
>>> *To:* serviceability-dev at openjdk.org
>>> *Subject:* Proposal to change the behavior of the timestamp place 
>>> holder (%t) in log file paths
>>>
>>> The timestamp place holder in a log filename currently expands to 
>>> the startup time of the JVM. When the log is rotated, the filename 
>>> containing this timestamp is suffixed with a file number. My 
>>> colleagues had expected the placeholder to be evaluated when the 
>>> current log file is rotated. They expected the name of each rotated 
>>> file to indicate the time the file was created. I think this 
>>> expectation is not unreasonable, so I wanted to discuss the 
>>> possibility of changing how/when the timestamp placeholder is 
>>> evaluated. If there is any appetite for a change like this, I am 
>>> willing to do the work. I would prefer to sort out the details 
>>> before coding anything, rather than discussing them in a surprise 
>>> pull request.
>>>
>>> Thank you for reading,
>>>
>>> William
>>>
>>
>