It's really fast; if it doesn't meet your requirements you're probably already aware that you need a specialized solution.
Besides explicit indexing, it also analyzes your data and maintains stats that inform the query planner. As your tables grow and your schema evolves, it will change its query plans accordingly. It can be quite smart. (A few examples: it maintains histograms of common values, null percentages, correlation between physical ordering and logical ordering or rows, and much more).
Of course, you also have a lot of flexibility to denormalize and get the best of both worlds.
But yes, generally speaking, I'm also surprised to see that in 2013 we're creating our indexes manually. While there are of course applications where you really want to ensure that all the right indexes are created ahead of time, for most applications having indexes created automatically based on query patterns would seem like a much better solution.
In a way, this is what caches can do. Besides that, Oracle is already very good at giving you statistics and tuning hints to help you assess where you could add an index:
http://stackoverflow.com/a/2937047/521799
DB2 and SQL Server probably have similar tools. As far as I know, they all don't go as far as automatically creating or dropping any indexes.
Many databases also support XML data types. And of course, you can always store unstructured data in CLOBs and BLOBs
What I was trying to imagine in my post, though, was some hypothetical DB where all this would be abstracted away--i.e., it would expose a NoSQL-like interface for all data, while working behind the scenes to provide SQL-like speed for frequent queries. Hence the analogy that it would be the "tracing JIT" of databases.
There are several reasons why you would not want to design a large-scale database this way. It implies an enormous amount of extra data motion which is arguably the major killer of performance in parallel databases; it is the same reason Hadoop is so slow and inefficient for analytic queries. Additionally, secondary indexing structures and similar types of de-normalization offer poor performance in distributed environments due to the implied consistency and coordination problems; it will severely degrade insert/update/delete operation performance as the data structures grow. Most database applications also value a high-degree of predictability of performance under load, which this definitely will not allow. These tradeoffs are unacceptable for many (most?) applications.
In short, while you could design a database that does what you suggest, it would only be useful at tiny scales or for read-only databases where consistent performance characteristics are not a requirement.
well, that is true for the whole SQL, but certainly not true for some subset of SQL which is rather big. And how much is "quite large"? We have client who claims that 160GB is big (he is running cluster of 8 machines now, poor bastard)
I normally design around databases in the tens of terabytes to tens of petabytes range, it is not an unusual scale. A database engine designed as the parent outlined would start to exhibit unacceptable performance characteristics at single-digit terabyte scales. A terabyte is tiny; I can buy servers with that much memory.
Complex multi-attribute selection and joins are efficiently parallelizable but most databases do not implement algorithms and data structures that allow this to be realized.
It seems maybe some of the NewSQL stuff is doing this. You use "SQL" but the data is still organized by some kind of hierarchy (like in F5). I'm trying to get into Google's Dremel paper to learn more about how they do it.
In general, though, it seems if there's any chance something might be normalized, but behind the scenes maybe its actually denormalized a bit, then folks try to layer on some kind of SQL-likeness to the database.
The solution in most cases is to maintain transaction and query optimized databases separately.
Not to stray too far off topic, but this article raised a question that I've had since starting software development in earnest... Why do we have to choose between heavily "normalized" relational databases that structure all data & thus allow more arbitrary queries, OR un-structured databases that are fast and flexible but often slow to query in complex ways? Why can't there be a hybrid "smart" database that dynamically generates indexes (to use the term loosely) based on how it's queried, in order to speed up similar queries in the future? With some kind of additional weight given to the most frequent queries. Granted, it would need some time to "warm up" (not unlike a tracing JIT compiler?), and the implementation might be fairly complex, but other than that, I can't think of any downsides...?