Hacker News Reader

Comment by elblanco | Hacker News Reader

This post just generated a tremendous amount of discussion in my company. Not just for the SQLite wrapper, but also about the Berkeley XML DB. Anyone have experience using Berkeley for medium-to-large scale data stores (1-100GB?)? How's the concurrency? What's the licensing model? Is there a way we can use this without having to give away our source?

View on HN

Project Voldemort uses BerkeleyDB Java Edition as one of the storage engines (the most frequently used one). Java Edition differs from the C version in being log-structured (very fast writes, at the cost of somewhat slower reads). In one deployment, we're able to store ~130 gb per node data (each node having 32 gb of memory, 18gb of it going to JVM heap, 10gb of the JVM heap going to BDB cache), while having average read latency times (which include the overhead of decoding a request packet) of <1 ms (with some outliers due to large value sizes, etc...).

I should add, this is also while using a single bdb environment for all partitions that belong to a node -- note the most efficient way to store this data (bdb je 4.0 does support sharing a cache between multiple environments, so we might move to separate environment per partition model). We've also built the distribution layer ourselves, rather than use the built-in HA features, but if I recall correctly, the first version of Amazon SimpleDB (prior to Erlang re-write) was built with BDB-JE HA (not to be confused with Amazon Dynamo, which is built on regular BDB-JE and provides its own distribution mechanisms).

Overall, we've found that BDB-JE performs best when there's 1:5 ratio between memory and data set size. With JE (as opposed to classic BDB) the issues are: Java implies a certain vertical scalability limit (due to GC pauses with very large heap sizes, however the GC does get better with every release), log-structured B+Tree is great for writes, but might mean more seeks being made for reads (especially if you were to use range query functionality).

Licensing might be an issue if you're shipping a commercial product to customer site. It doesn't matter if it's used on your own servers or in an open source project (and unlike GPL, the BerkeleyDB license is compatible with Apache licensed projects). The licensing also doesn't apply to Perl/Ruby/Python/other dynamic language bindings for BerkeleyDB, you're free to package them (and then request the customer to install BerkeleyDB themselves, much like you would with MySQL).

In short, it's a great solution if you fit its use case. I should also add that interacting with the Sleepycat people online (e.g., on the forums) also makes you forget they're a part of Oracle.

I don't have experience with the Berkeley XML DB. I have used the BerkeleyDB 'Java Edition' (a different codebase and on-disk format) for multi-GB spill-to-disk queues and maps/sets.

The BerkeleyDB license is roughly: free to include in open-source projects, pay if you want to redistribute as part of a closed-source product. So like with the GPL, using strictly inside your company might not be considered 'distribution'. But also, if you are selling software licenses their bundling rates might not be bad compared to your unit price. Of course, do your own license analysis before business use.

What would be the licensing model if the product is open source, but is NOT free (i.e., customers are charged for an open-source product in which BDB is used)?

You need to do your own analysis, not trusting some random guy on the internet (eg, me). But:

(1) If people need to pay to use, it's not "open source" by the "open source definition" of the OSI: <http://www.opensource.org/docs/osd>;

(2) Contacting Oracle/BDB will let you know what they think your obligations are.

Reply on news.ycombinator.com