Libbson is a new shared library written in C for developers wanting to work with the BSON serialization format.
Its API will feel natural to C programmers but can also be used as the base of a C extension in higher-level MongoDB drivers.
The library contains everything you would expect from a BSON implementation. It has the ability to work with documents in their serialized form, iterating elements within a document, overwriting fields in place, Object Id generation, JSON conversion, data validation, and more. Some lessons were learned along the way that are beneficial for those choosing to implement BSON themselves.
A common use case of BSON is for relatively small documents. This has a profound impact on the memory allocator in userspace, causing what is commonly known as “memory fragmentation”. Memory fragmentation can make it more difficult for your allocator to locate a contiguous region of memory.
In addition to increasing allocation latency, it increases the memory requirements of your application to overcome that fragmentation.
To help with this issue, the bson_t structure contains 120 bytes of inline space that allows BSON documents to be built directly on the stack as opposed to the heap.
When the document size grows past 120 bytes it will automatically migrate to a heap allocation.
Additionally, bson_t will grow it’s buffers in powers of two. This is standard when working with buffers and arrays as it amortizes the overhead of growing the buffer versus calling realloc() every time data is appended. 120 bytes was chosen to align bson_t to the size of two sequential cachelines on x86_64 (each 64 bytes).
This may change based on future research, but not before a stable ABI has been reached.
One strength of BSON is it’s ability to nest objects and arrays. Often times when serializing these nested documents, each sub-document is serialized independently and then appended to the parents buffer.
As you might imagine, this takes quite the toll on the allocator. It can generate many small allocations which were only created to have been immediately discarded after appending to the parents buffer. Libbson allows for building sub-documents directly into the parent documents buffer.
Doing so helps avoid this costly fragmentation. The topmost document will grow its underlying buffers in powers of two each time the allocation would overflow.
Another common area for allocator fragmentation is during BSON document parsing. Libbson allows parsing and iteration of BSON documents directly from your incoming network buffer.
This means the only allocations created are those needed for your higher level language such as a PyDict if writing a Python extension.
Developers writing C extensions for their driver may choose to implement a “generator” style parsing of documents to help keep memory fragmentation low.
A technique we’re yet to explore is implementing a hashtable-esque structure backed by BSON, only deserializing the entire buffer after a threshold of keys have been accessed.
Much like parsing BSON documents, generating documents and placing them into your network buffers can be hard on your memory allocator. To help keep this fragmentation down, Libbson provides support for serializing your document to BSON directly within a buffer of your choosing.
This is ideal for situations such as writing a sequence of BSON documents into a MongoDB message.
Applications are often doing ObjectId generation, especially in high insert environments. The uniqueness of generated ObjectIds is critical to avoiding duplicate key errors across multiple nodes.
Highly threaded environments create a local contention point slowing the rate of generation. This is because the threads must synchronize on the increment counter of each sequential ObjectId. Failure to do so could cause collisions that would not be detected until after a network round-trip. Most drivers implement the synchronization with an atomic increment or a mutex if atomics are not available.
Libbson will use atomic increments and in some cases avoid synchronization altogether if possible. One such case is a non-threaded environment.
Another is when running on Linux as both threads and processes are in the same namespace.
This allows the use of the thread identifier as the pid within the ObjectId.