Active Data Vs. Active ArchiveWe need better metrics to help us decide what data should be on primary storage and what should be on archive storage.
In my last column I discussed how what we used to consider active data is changing. We now have to look at the potential working set instead of the actual working set. Thanks to initiatives like real-time analytics, some data that we used to classify as archivable now needs to be at the ready. If this is the case, what is the role of archive? How do disk and tape archives participate in an increasingly active world?
The key to a balanced storage strategy, even with all this active data, is to change how we decide to archive a certain set of data. Under the current archive methodology the most common decision point was last modification date. In other words, data that is X days/years old can be archived, everything else has to stay on primary storage. The problem with this methodology is it is not compatible with real-time analytics and not even really compatible with the way users use data.
We need better metrics to help us decide what data should be on primary storage and what should be on archive storage. A key criteria is going to be what data, if it needs to be accessed, will need to be delivered instantly -- in other words, something that may need to be analyzed in the future. This data should probably not go to an archive no matter how old it gets since it could have a statistical probability of value.
[ Learn more about virtual desktop infrastructure. Read VDI Performance And Cost: A Deeper Dive. ]
However, if we know for sure that a certain data set will not be part of a real-time processing application or be needed for analytics then lets archive it as soon as possible and not even wait for it to age. Maybe some of this data could even spend all of its data lifecycle on archive storage because the performance of the archive is "good enough" for the use case.
There is also the need to understand relationships between files. As a simple example, I am writing a couple of books right now. Each of those books have multiple iterations on the file name but large chunks of the content within those files are the same. Each draft gets a different file name. When I get to the end of any of these books, I really don't think I will need all of these drafts but, because all data has become a "you never know" situation, I will want to keep all of them around but I doubt I will ever access them again.
The question is how many of these drafts will I require instant access to and how many could I wait 10 minutes before I view them? For my purposes, all I really will need is the final copy and maybe a couple of the iterations. It would be nice to have software analyze this data and keep versions of the files with the most significant internal changes and then archive the rest.
Interestingly, one of the things we are learning from our primary storage deduplication test is how big of a role this technology can play in these circumstances. Essentially, I can keep all of the files with minimal impact on space utilization. And since they can be disk-based, retrieval time is excellent.
Another classification point is how is that data acted on when recovered? From beginning to end, or at some random point in the file? Basically, can the data be utilized sequentially? If this is the case, then just the front section of that data needs to be stored on primary storage, enough so that it can start being accessed while the back end catches up and the users see no delay in response time. This capability will require a file system intelligent enough to deliver data from two different sources at the same time.
When these attributes of the data are known and understood, then it can be properly placed in the proper types of storage in the data center. Data that whose recovery need is random and unpredictable will need to go on fast storage if analytics are being used. Data that is very similar to other data can be archived or deduplicated.
This archive, depending on what the known recovery need is, can easily be tape based because for a large chunk of the data set how quickly it is recovered is less import than how cost effectively can it be stored.