Pgfs: The PostGres File System

The details of how Pgfs came to be written and how it can save you disk space.
Audience Participation Time

So far I've identified versets only by integers, but integers are boring. Since Pgfs is built on top of a database, all manner of things are possible. What naming schemes have you come up with for versets in other projects to support a configuration management effort? Are all the names/identifiers in a flat space? Are they hierarchical, and do they inherit properties from the location they are placed in? Let me know, I'm particularly interested in experiences from large projects with tens and hundreds of thousands of files per verset.

So far the only method I've supplied for making a new verset is by sending Pgfs a special command to copy one verset into a new one. However, a verset is just a collection of database rows, so it can be manufactured by a SQL program, perhaps one that represents a semi-automated multi-way merge across 14 versets. Take the base file tree from here, take this patch there, take this other patch over yonder, and make all the daemons owned by fred, thank you very much. What would an interface to control this process look like? Would you have an interactive file-browser shopping-cart thing, where you pulled bits and pieces from wherever you find them? How would this process resolve collisions?

There are more interesting open-ended questions in the BUGS and TODO files of the Pgfs distribution that concern both interface and implementation. I encourage you to pick a couple that interest you and talk about them on the host-gen mailing list.

Closing Thoughts

The most important message I want to give you is that file system hacking is not just for wizards anymore. NFS supplies a portable file system interface that eliminates the usual kernel-hacking requirements. NFS semantics are not great, but they are adequate for many things. Anyone can use the NFS-decoding portion of Pgfs as a skeleton and write a file system with whatever semantics they dream up. Mundane possibilities are an ftp-browsing or web-browsing file system. More interesting areas involve wide-area, fault-tolerant file systems with distributed physical redundancy. The job of higher-level protocols is to turn failure into bad performance. Instead of a list of Linux ftp sites to pick through by hand, wouldn't you rather use a file system that automatically gets blocks from the best-performing site of the moment? Wouldn't you like to create a new local storage area for a subtree of your favorite archive site, and the only thing your users need to know is...access just got faster? These ideas raise lots of interesting authentication and trust issues, many of which can be solved by the PGP model of the web-of-trust. Now, go forth and code.

Sidebar 3

Brian Bartholomew is writing Pgfs as a component of the Host Factory automated host maintenance system. Host Factory integrates hosts into a Borg collective. Working Version does large site toolsmithing, and further info on Host Factory is available at Brian can be reached at