Data structure/algorithm for variable length record storage and lookup on disk withsearch only on primary keys

Question

I am looking for an algorithm / data structure that works well for large block based devices (eg a mechanical hard drive) which is optimised for insert, get, update and delete where searches are always done using the id of the data and where the data fields for any ID have a variable length.

The B-Tree seems to be a commonly quoted structure but mainly for fixed length records. I also expect dramatically more gets and updates than I do inserts and deletes. Can I get rid of the O(log m) lookup of the B-tree?

I am quite happy for it to be a combined system, for instance ISAM combines a B-tree and linear file storage which looks like it can be made to work with variable length records as an approach. Is there something better?

Some further constraints:

1) IDs are potentially sparse but they can be made to come in blocks of linear numbers - but in a large range (64 bit)

2) I don't want to use a DBMS, performance for my particular problem hasn't proved very good. I don't need any of the operations a full DBMS uses, I don't need search. I need something I can tweak and optimise easily. Call it academic curiosity, if it is out performed by MySQL then I'll use that but I have to try and go faster.

3) The dataset is larger than can fit in memory, the index however may well fit into memory if its as simple as key, offset. I am certainly looking at something like 1 billion entities or more in storage.

4) Ideally space should be recovered when a record is deleted. That may be via compaction but I am interested to see if there is a better way (A B-tree for instance recovers space easily).

Answer 1

The easy way: Use something like Berkeley DB. It provides a key-value store for arbitrary byte strings, and does all the hard work for you. It even provides 'secondary databases' for indexing, if you want it.

The do-it-yourself way: Use Protocol Buffers (or the binary format of your choice) to define B-Tree node and data item structures. Use an append-only file for your database. To write a new record or modify an existing record, you simply write the record itself to the end of the file, then write any modified B-Tree nodes (eg, the record's parent node, its parent node, and so forth up to the root). Then, write the location of the new root of the tree to the header block at the beginning of the file. To read the file, you simply find the most recent root node and read the B-Tree like you would in any other file. This approach has several advantages:

• Since written data is never modified, readers don't need to take locks, and get a 'snapshot' view of the DB based on the root node at the time they started reading.
• By adding 'previous version' fields to your nodes and records, you get the ability to access previous versions of the DB essentially for free.
• It's really easy to implement and debug compared to most on-disk file formats that support modification.
• Compacting the database consists of simply reading out the latest version of the data and B-Tree and writing it to a new file.