在线数据存储系统S3QL

jieforest

Data Consistency

In contrast to the typical hard disk, most storage providers do not guarantee immediate consistency of written data. This means that:

1 after an object has been stored, requests to read this object may still fail or return the prior contents for a little while.

2 after an object has been deleted, attempts to read it may still return the (old) data for some time, and it may still remain in the list of stored objects for some time.

3 after a new object has been created, it may still not be included when retrieving the list of stored objects for some time.

Of course, none of this is acceptable for a file system, and S3QL generally handles any of the above situations internally so that it always provides a fully consistent file system to the user. However, there are some situations where an S3QL user nevertheless needs to be aware of the peculiarities of his chosen storage service.

Suppose that you mount the file system, store some new data, delete some old data and unmount it. If you then mount the file system again right away on another computer, there is no guarantee that S3QL will see any of the changes that the first S3QL process has made. At least in theory it is therefore possible that on the second mount, S3QL does not see any of the changes that you have done and presents you an “old version” of the file system without them. Even worse, if you notice the problem and unmount the file system, S3QL will upload the old status (which S3QL necessarily has to consider as current) and thereby permanently override the newer version (even though this change may not become immediately visible either). S3QL uses several techniques to reduce the likelihood of this to happen (see Implementation Details for more information on this), but without support from the storage service, the possibility cannot be eliminated completely.

jieforest

The same problem of course also applies when checking the file system. If the storage service provides S3QL with only partially updated data, S3QL has no way to find out if this a real consistency problem that needs to be fixed or if it is only a temporary problem that will resolve itself automatically (because there are still changes that have not become visible yet).

This is where the so called consistency window comes in. The consistency window is the maximum time (after writing or deleting the object) for which any of the above “outdated responses” may be received. If the consistency window is zero, i.e. all changes are immediately effective, the storage service is said to have immediate consistency. If the window is infinite, i.e. there is no upper bound on the time it may take for changes to become effect, the storage service is said to be eventually consistent. Note that often there are different consistency windows for the different operations. For example, Google Storage offers immediate consistency when reading data, but only eventual consistency for the list of stored objects.

To prevent the problem of S3QL working with an outdated copy of the file system data, it is therefore sufficient to simply wait for the consistency window to pass before mounting the file system again (or running a file system check). The length of the consistency window changes from storage service to storage service, and if your service is not included in the list below, you should check the web page or ask the technical support of your storage provider. The window that is important for S3QL is the smaller of the times for which

1 a newly created object may not yet be included in the list of stored objects

2 an attempt to read a newly created object may fail with the storage service reporting that the object does not exist

Unfortunately, many storage providers are hesitant to guarantee anything but eventual consistency, i.e. the length of the consistency window is potentially infinite. In that case you simply have to pick a length that you consider “safe enough”. For example, even though Amazon is only guaranteeing eventual consistency, the ordinary consistency window for data stored in S3 is just a few seconds, and only in exceptional circumstances (i.e., core network outages) it may rise up to hours (source).

jieforest

Data Durability

The durability of a storage service a measure of the average probability of a storage object to become corrupted over time. The lower the chance of data loss, the higher the durability. Storage services like Amazon S3 claim to achieve a durability of up to 99.999999999% over a year, i.e. if you store 100000000 objects for 100 years, you can expect that at the end of that time one object will be corrupted or lost.

S3QL is designed to reduce redundancy and store data in the smallest possible form. Therefore, S3QL is generally not able to compensate for any such losses, and when choosing a storage service you should carefully review if the offered durability matches your requirements. When doing this, there are two factors that should be kept in mind.

Firstly, even though S3QL is not able to compensate for storage service failures, it is able to detect them: when trying to access data that has been lost or corrupted by the storage service, an IO error will be returned and the mount point will become inaccessible to ensure that the problem is noticed.

Secondly, the consequences of a data loss by the storage service can be significantly more severe than you may expect because of S3QL’s data de-duplication feature: a data loss in the storage service at time x may cause data that is written after time x to be lost as well. Consider the following scenario:

1 You store an important file in the S3QL file system.

2 The storage service loses the data blocks of this file. As long as you do not access the file or run fsck.s3ql, S3QL is not aware that the data has been lost by the storage service.

3 You save an additional copy of the important file in a different location on the same S3QL file system.

4 S3QL detects that the contents of the new file are identical to the data blocks that have been stored earlier. Since at this point S3QL is not aware that these blocks have been lost by the storage service, it does not save another copy of the file contents in the storage service but relies on the (presumably) existing blocks instead.

5 Therefore, even though you saved another copy, you still do not have a backup of the important file (since both copies refer to the same data blocks that have been lost by the storage service).

jieforest

For some storage services, fsck.s3ql can mitigate this effect. When fsck.s3ql runs, it asks the storage service for a list of all stored objects. If objects are missing, it can then mark the damaged files and prevent the problem from spreading forwards in time. Figuratively speaking, this establishes a “checkpoint”: data loss that occurred before running fsck.s3ql can not affect any file system operations that are performed after the check. Unfortunately, many storage services only “discover” that objects are missing or broken when the object actually needs to be retrieved. In this case, fsck.s3ql will not learn anything by just querying the list of objects.

In the future, fsck.s3ql will have an additional “full-check” mode, in which it attempts to retrieve every single object. However, this is expected to be rather time consuming and expensive. Therefore, it is generally a better choice to choose a storage service where the expected data durability is so high that the possibility of a lost object (and thus the need to run any full checks) can be neglected over long periods of time.

To some degree, fsck.s3ql can mitigate this effect. When used with the --full-check option, fsck.s3ql asks the storage service to look up every stored object. This way, S3QL learns about any missing and, depending on the storage service, corrupted objects. It can then mark the damaged files and prevent the problem from spreading forwards in time. Figuratively speaking, this establishes a “checkpoint”: data loss that occurred before running fsck.s3ql with --full-check can not affect any file system operations that are performed after the check.

Unfortunately, a full check is rather time consuming and expensive because of the need to check every single stored object. It is generally a better choice to choose a storage service where the expected data durability is so high that the possibility of a lost object (and thus the need to run any full checks) can be neglected over long periods of time.

jieforest

over.