The physical database size doesn't matter. The number of records don't matter.
In my experience the biggest problem that you are going to run in to is not size, but the number of queries you can handle at a time. Most likely you are going to have to move to a master/slave configuration so that the read queries can run against the slaves and the write queries run against the master. However if you are not ready for this yet, you can always tweak your indexes for the queries you are running to speed up the response times. Also there is a lot of tweaking you can do to the network stack and kernel in Linux that will help.
I have had mine get up to 10GB, with only a moderate number of connections and it handled the requests just fine.
I would focus first on your indexes, then have a server admin look at your OS, and if all that doesn't help it might be time to implement a master/slave configuration.
My favorite answer is as what the first sentence in this thread suggested. Use an Adjacency List to maintain the hierarchy and use Nested Sets to query the hierarchy.
The problem up until now has been that the coversion method from an Adjacecy List to Nested Sets has been frightfully slow because most people use the extreme RBAR method known as a "Push Stack" to do the conversion and has been considered to be way to expensive to reach the Nirvana of the simplicity of maintenance by the Adjacency List and the awesome performance of Nested Sets. As a result, most people end up having to settle for one or the other especially if there are more than, say, a lousy 100,000 nodes or so. Using the push stack method can take a whole day to do the conversion on what MLM'ers would consider to be a small million node hierarchy.
I thought I'd give Celko a bit of competition by coming up with a method to convert an Adjacency List to Nested sets at speeds that just seem impossible. Here's the performance of the push stack method on my i5 laptop.
Duration for 1,000 Nodes = 00:00:00:870
Duration for 10,000 Nodes = 00:01:01:783 (70 times slower instead of just 10)
Duration for 100,000 Nodes = 00:49:59:730 (3,446 times slower instead of just 100)
Duration for 1,000,000 Nodes = 'Didn't even try this'
And here's the duration for the new method (with the push stack method in parenthesis).
Duration for 1,000 Nodes = 00:00:00:053 (compared to 00:00:00:870)
Duration for 10,000 Nodes = 00:00:00:323 (compared to 00:01:01:783)
Duration for 100,000 Nodes = 00:00:03:867 (compared to 00:49:59:730)
Duration for 1,000,000 Nodes = 00:00:54:283 (compared to something like 2 days!!!)
Yes, that's correct. 1 million nodes converted in less than a minute and 100,000 nodes in under 4 seconds.
You can read about the new method and get a copy of the code at the following URL.
http://www.sqlservercentral.com/articles/Hierarchy/94040/
I also developed a "pre-aggregated" hierarchy using similar methods. MLM'ers and people making bills of materials will be particularly interested in this article.
http://www.sqlservercentral.com/articles/T-SQL/94570/
If you do stop by to take a look at either article, jump into the "Join the discussion" link and let me know what you think.
Best Answer
Adding to @Oscar-Godson's excellent answer, you're likely to experience problems with values much larger than 50MBs. Bitcask is best suited for values that are up to a few KBs. If you're storing large values, you may want to consider alternative storage backends, such as innostore.
I don't have experience with storing binary values, but we've a medium-sized cluster in production (5 nodes, on the order of 100M values, 10's of TBs) and we're seeing frequent errors related to inserting and retrieving values that are 100's of KBs in size. Performance in this case is inconsistent - some times it works, others it doesn't - so if you're going to test, test at scale.
We're also seeing problems with large values when running map-reduce queries - they simply time out. However that may be less relevant to binary values... (as @Matt-Ranney mentioned).
Also see @Stephen-C's answer here