Patch to handle missing /etc/localtime

[rickysarraf]

On Arch Linux, the documentation clearly states that /etc/localtime is not required, and in CI it's not present. Patch the timezone handling to assume UTC when /etc/localtime is not there.

This is hopefully a sane fallback for systems where /etc/localtime may not be available.

--- a/c++/src/Timezone.cc             2025-03-12 12:09:01.267309399 +0000
+++ b/c++/src/Timezone.cc             2025-03-12 12:12:12.303295722 +0000
@@ -28,6 +28,12 @@
 #include <map>
 #include <sstream>
 
+#ifndef _MSC_VER
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+#endif
+
 namespace orc {
 
   // default location of the timezone files
@@ -694,6 +700,14 @@
 #ifdef _MSC_VER
     return getTimezoneByName("UTC");
 #else
+    // The absence of LOCAL_TIMEZONE means UTC conventionally
+    {
+      struct stat _ignored;
+      if (stat(LOCAL_TIMEZONE, &_ignored) == -1) {
+        return getTimezoneByName("UTC");
+      }
+    }
+
     return getTimezoneByFilename(LOCAL_TIMEZONE);
 #endif
   }

[dongjoon-hyun]

Thank you for reporting, @rickysarraf .

cc @wgtmac

[wgtmac]

I'm not sure if such fallback makes sense because it may slightly produce data shift for timestamp values written by a different host. I also see that the fallback exists on Windows.

@ffacs WDYT?

[ffacs]

I'm not sure if such fallback makes sense because it may slightly produce data shift for timestamp values written by a different host. I also see that the fallback exists on Windows.

@ffacs WDYT?

I think the fallback makes sense. Because when using the local timezone, reading on different machines may produce different results.

[wgtmac]

The problem may come from the writer which blindly uses UTC to serialize timestamp values. In terms of fallback, I think it is better to make the fallback value configurable and fail if unset.

[dongjoon-hyun]

Let me close this because the claim is not true. The typical example is the old timezone ID, US/Pacific which is currently replaced into America/Los_Angeles in the new naming scheme. Currently, Java and some OSes supports this kind of legacy identifiers as a fallback, but new OSes have only new timezone IDs in these days.

• https://data.iana.org/time-zones/tzdb/backward

The absence of LOCAL_TIMEZONE means UTC conventionally

[lahwaacz]

Let me close this because the claim is not true. The typical example is the old timezone ID, US/Pacific which is currently replaced into America/Los_Angeles in the new naming scheme. Currently, Java and some OSes supports this kind of legacy identifiers as a fallback, but new OSes have only new timezone IDs in these days.

* https://data.iana.org/time-zones/tzdb/backward

There is a difference between /etc/localtime being a symlink to a non-existent file (which is what you described) and the /etc/localtime file missing completely (which is the situation in Arch Linux containers). What if the proposed fallback was changed to use lstat instead of stat?

[dongjoon-hyun]

Even in that case, I don't think there is a common consensus that the missing file assumes UTC. Do you have any POSIX documentation link to support your claim?

There is a difference between /etc/localtime being a symlink to a non-existent file (which is what you described) and the /etc/localtime file missing completely (which is the situation in Arch Linux containers)

[lahwaacz]

Do you have any POSIX documentation link that supports yours?

There is a localtime(5) man page (part of systemd) which says:

If /etc/localtime is missing, the default "UTC" timezone is used.

[dongjoon-hyun]

Oh, my bad. Thank you for pointing out that, @lahwaacz .

[dongjoon-hyun]

I reopened. Is it only for Arch linux? What about other OSes?

[lahwaacz]

systemd is not an Arch Linux project 🤷

[dongjoon-hyun]

Thanks. I confirmed that it's universal.

[dongjoon-hyun]

Could you make a PR with this, @lahwaacz ? That seems to be the only remaining concerns on this discuss thread.