I've built a web service in PHP for answering the requests from the clients. The parsed data is being stored in SQLite databases on SSD. I use multiple databases to be able to re-parse only some parts of the blockchain when needed (by copying the files to a backup path on my server, so that the clients can still retrieve the old data in the meantime. With SQLite it's also possible to join data from multiple DBs in one SQL statement).

I currently don't run an own node on that server and the new blocks are being downloaded from other public instances with cURL (I will start an own one in the coming months) and they are parsed each minute (with help of Cron). The blocks are being stored in SQLite databases (100,000 blocks per file), which (when full) are being compressed with XZ, which makes them much smaller in size. To save even more SSD space I'm using an own algorithm for JSON compression.

If you want to access the account history in real time, you would need to start an own node with enabled account history plugin and parse the virtual ops of the incoming blocks yourself (or use the inbuilt RocksDB solution). I currently would not recommend this and it may be better to use Hivemind for that, but I guess you will nevertheless have the same distance (a few minutes) to the current head block in that case.