From: Daira Hopwood Date: Wed, 27 May 2015 11:19:36 +0000 (+0100) Subject: Describe the procedure for writes that are initially classified as conflicts. X-Git-Url: https://git.rkrishnan.org/reliability?a=commitdiff_plain;h=68adf5b41b8a2f10e0fd6a04029252e1a8e68074;p=tahoe-lafs%2Ftahoe-lafs.git Describe the procedure for writes that are initially classified as conflicts. Signed-off-by: Daira Hopwood --- diff --git a/docs/proposed/magic-folder/remote-to-local-sync.rst b/docs/proposed/magic-folder/remote-to-local-sync.rst index eaec05f6..2937be73 100644 --- a/docs/proposed/magic-folder/remote-to-local-sync.rst +++ b/docs/proposed/magic-folder/remote-to-local-sync.rst @@ -432,7 +432,7 @@ Therefore we need to consider carefully how to handle failure conditions. In our proposed design, Alice's Magic Folder client follows -this procedure for each download in response to a remote change: +this procedure for an overwrite in response to a remote change: 1. Write a temporary file, say ``.foo.tmp``. 2. If there are pending notifications of changes to ``foo``, @@ -641,6 +641,18 @@ open will fail because ``foo`` does not exist. Nevertheless, no data will be lost, and in many cases the user will be able to retry the operation. +Above we only described the case where the download was initially +classified as an overwrite. If it was classed as a conflict, the +procedure is the same except that we choose a unique filename +for the conflicted file (say, ``foo.conflicted_unique``). We write +the new contents to ``.foo.tmp`` and then rename it to +``foo.conflicted_unique`` in such a way that the rename will fail +if the destination already exists. (On Windows this is a simple +rename; on Unix it can be implemented as a link operation followed +by an unlink, similar to steps 4c and 4d above.) If this fails +because another process wrote ``foo.conflicted_unique`` after we +chose the filename, then we retry with a different filename. + Read/download collisions ~~~~~~~~~~~~~~~~~~~~~~~~