On Tue, Sep 02, 2014 at 01:56:34AM -0400, Tom Lane wrote:
> Noah Misch <noah@leadboat.com> writes:
> > On Mon, Sep 01, 2014 at 10:22:57PM -0400, Tom Lane wrote:
> >> Also, as best I can tell, .psql_history files from older libedit versions
> >> are not forward-compatible to current libedit versions because of the
> >> failure of the decode_history() loop to reach all lines of the file
> >> when using current libedit. That is also a back-patchable bug fix IMO.
> >> (Closer investigation suggests this is a bug or definitional change in
> >> libedit's history_set_pos, not so much in next_history vs
> >> previous_history. But whatever it is, it behooves us to work around it.)
>
> > I haven't studied this part of the topic other than to read what you have
> > written. All other things being equal, I agree. If fixing this will make
> > psql-9.3.6 w/ libedit-20141001 write history files that confuse psql-9.3.5 w/
> > libedit-20141001, that changes the calculus. Will it?
>
> I'm not sure exactly when things changed, but I have verified that the
> existing loops in decode/encode_history visit all lines of the history
> when using OS X Tiger's libedit library. On OS X Mavericks, the loops
> visit only the oldest history entry, as Stepan reported. This means that
> there may be libedit-style ~/.psql_history files out there in which ^A has
> been substituted for ^J (in lines after the oldest), which will not be
> correctly reloaded by psql versions using newer libedit.
>
> It's certainly arguable whether this is an issue warranting a back-patch,
> since we've not heard field complaints about it AFAIR. But I think we
> ought to do so. I think "psql N produces files that psql N+1 can't read"
> is worse than the reverse case, and that's exactly what we're debating
> here.
I tried your patches against libedit-28. Wherever a command contains a
newline, unpatched psql writes the three bytes "\^A" to the history file, and
patched psql writes the four bytes "\012". Unpatched psql correctly reads
either form of the history file. Patched psql misinterprets a history file
created by unpatched psql, placing 0x01 bytes in the recalled command where it
should have newlines. That's a worrisome compatibility break.