Servers of Happiness

When you upload a file to a Tahoe-LAFS grid, you expect that it will stay there for a while, and that it will do so even if a few of the peers on the grid stop working, or if something else goes wrong. An upload health metric helps to make sure that this actually happens. An upload health metric is a test that looks at a file on a Tahoe-LAFS grid and says whether or not that file is healthy; that is, whether it is distributed on the grid in such a way as to ensure that it will probably survive in good enough shape to be recoverable, even if a few things go wrong between the time of the test and the time that it is recovered. Our current upload health metric for immutable files is called ‘servers-of-happiness’; its predecessor was called ‘shares-of-happiness’.

shares-of-happiness used the number of encoded shares generated by a file upload to say whether or not it was healthy. If there were more shares than a user-configurable threshold, the file was reported to be healthy; otherwise, it was reported to be unhealthy. In normal situations, the upload process would distribute shares fairly evenly over the peers in the grid, and in that case shares-of-happiness worked fine. However, because it only considered the number of shares, and not where they were on the grid, it could not detect situations where a file was unhealthy because most or all of the shares generated from the file were stored on one or two peers.

servers-of-happiness addresses this by extending the share-focused upload health metric to also consider the location of the shares on grid. servers-of-happiness looks at the mapping of peers to the shares that they hold, and compares the cardinality of the largest happy subset of those to a user-configurable threshold. A happy subset of peers has the property that any k (where k is as in k-of-n encoding) peers within the subset can reconstruct the source file. This definition of file health provides a stronger assurance of file availability over time; with 3-of-10 encoding, and happy=7, a healthy file is still guaranteed to be available even if 4 peers fail.

Measuring Servers of Happiness

We calculate servers-of-happiness by computing a matching on a bipartite graph that is related to the layout of shares on the grid. One set of vertices is the peers on the grid, and one set of vertices is the shares. An edge connects a peer and a share if the peer will (or does, for existing shares) hold the share. The size of the maximum matching on this graph is the size of the largest happy peer set that exists for the upload.

First, note that a bipartite matching of size n corresponds to a happy subset of size n. This is because a bipartite matching of size n implies that there are n peers such that each peer holds a share that no other peer holds. Then any k of those peers collectively hold k distinct shares, and can restore the file.

A bipartite matching of size n is not necessary for a happy subset of size n, however (so it is not correct to say that the size of the maximum matching on this graph is the size of the largest happy subset of peers that exists for the upload). For example, consider a file with k = 3, and suppose that each peer has all three of those pieces. Then, since any peer from the original upload can restore the file, if there are 10 peers holding shares, and the happiness threshold is 7, the upload should be declared happy, because there is a happy subset of size 10, and 10 > 7. However, since a maximum matching on the bipartite graph related to this layout has only 3 edges, Tahoe-LAFS declares the upload unhealthy. Though it is not unhealthy, a share layout like this example is inefficient; for k = 3, and if there are n peers, it corresponds to an expansion factor of 10x. Layouts that are declared healthy by the bipartite graph matching approach have the property that they correspond to uploads that are either already relatively efficient in their utilization of space, or can be made to be so by deleting shares; and that place all of the shares that they generate, enabling redistribution of shares later without having to re-encode the file. Also, it is computationally reasonable to compute a maximum matching in a bipartite graph, and there are well-studied algorithms to do that.

Issues

The uploader is good at detecting unhealthy upload layouts, but it doesn’t always know how to make an unhealthy upload into a healthy upload if it is possible to do so; it attempts to redistribute shares to achieve happiness, but only in certain circumstances. The redistribution algorithm isn’t optimal, either, so even in these cases it will not always find a happy layout if one can be arrived at through redistribution. We are investigating improvements to address these issues.

We don’t use servers-of-happiness for mutable files yet; this fix will likely come in Tahoe-LAFS version 1.13.