So... this NoSQL thing

I am also a big fan of MongoDB. That having been said, it is absolutely not a wholesale replacement for RDBMS. Facebook has 300 million users but if some of your friends don't show up in the list one time, or one of the photo albums is missing on the occasional request, would you notice? Probably not. If your status update doesn't trickle down to all of your friends for a few minutes, does it matter? Hardly. If Wal-Mart's balance sheets are out of sync, would someone lose their head? Definitely.

NoSQL databases are great in "fuzzy" environments where relationships are not strict and data integrity can afford to be out of sync. RDBMS are still important when data sets are extremely complex and relational (hence the name), and they need to be kept pure.

The big push to NoSQL comes from the fact for the last 30 years, we have been using RDMBS systems for both scenarios. We now have a more appropriate tool for many situations. Some would argue most, in fact. But no one would argue all.

I write this but as a dispute to Rex's answer.

I dispute the idea that nosql is relationless and fuzzy.

I had been working with CODASYL many years ago with C and Cobol - entity relationships are very tight in CODASYL.

In contrast, relational database systems have a very liberal policy towards relationships. As long as you can identiy a foreign key, you could form a relationship adhoc.

It is frequently taken for granted that SQL is synonymous with RDBMS, but people have been writing SQL drivers for CODASYL, XML, inverted sets, etc.

RDBMS/SQL do not equal precision in data or relationship. In fact, RDBMS has been a constant cause in imprecision and misperception of relationships. I do not see how RDBMS offer better data and relationship integrity than hadoop, for example. Put on a layer of JDO - and we can construct a network of good and clean relationships between entities in hadoop.

However, I like working with SQL because it gives me the ability to script adhoc relationships, even though I realise that adhoc relationships is a constant cause of relationship adulteration and problems.

Having the opportunity to work with statistical analysis of business and industrial processes, SQL gave me the ability to explore relationships where no relationships had previously been perceived. The opportunity to work with statistical analysis gave me insights that would not normally come the way of SQL programmers.

For example, you would design and normalise your schema to reflect a set of processes. What you might not realise is that relationships change over time. The statistical characteristics would reveal that a schema may no longer be as "properly normalised" as it once had been. That the principal components of the processes have mutated over time. But non-statistical programmers do not understand that and continue to tout RDBMS as the perfect solution for data integrity and relationship precision.

However, in a relationship-linking database, you could link entities in relationships as they appear. When relationships mutate, the linking naturally mutate with the data. Relationships and their mutation are documented within the database system without the expensive need to renormalise the schema. At which point, RDBMS is good only as temp dbs.

But then you might counter that RDBMS too allows you to flexibly mutate your relationships, since that is what SQL does best. True, very true - so long as you perform BCNF or even 4NF. Otherwise, you would begin to see that your queries and data loaders performing replicated operations. But then your many years in the RDBMS business have so far certainly at least made you realise that BCNF is very expensive and operationally inefficient and that we are constantly guilty of 2.5 NFing our schemata.

To say that RDBMS and SQL promotes data and relationship integrity is a gross mis-statement. Either you work in a company that is so tiny or you didn't stay in your positions for more than two years - you would not see the amount of data or the information mutation and the problems caused by RDBMS. The abuse of RDBMS is the cause of executives being restricted in the view by computer applications and the cause of financial failures of companies failing to see changes in market behaviour because their views were restricted by the programmers whose views were restricted to their veneration of their beloved RDBMS schemata.

That is why SQL programmers do not understand why your company statistician refuses to use your application which you crafted meticulously but they employed a college intern to write SQL to download data into their personal servers and that your company executives learn to trust the accountants' and statisticians' spreadsheets rather than your elegant multi-tiered applications because of your applications' inability to mutate with processes.

It might not be possible, but I still urge you to acquire some statistical understanding to perceive how processes mutate over time so that you can make the right technological decision.

The reason people are not moving to SQL-less is lack of a good scripting environment like SQL to perform adhoc relationship analysis. Not because SQL-less technology is deficient in precision or integrity. Adhoc relationship analysis is very important nowadays due to the rapid and agile application development attitudes and strategies we have nowadays.

Let me hit the questions one at a time:

I know I can't do transactions across relationships... when would this be a big deal?

Picture cascading deletes. Or even just basic referential integrity. The concept of "foreign keys" can't really be enforced across "collections" (the Mongo term for tables). You can do atomic writes to only a single "document" (AKA record). So if you have a DB issue, you can orphan data in the DB.

I get as much performance as I have CPUs and RAM for free?

Not free, but definitely with a different set of trade-offs. For example, Mongo is great at running single-record, key/value look-ups. However, Mongo is poor at running relational queries. You'll need to use map-reduce for many of these. Mongo is a "RAM-whore". Mongo basically demands 64-bit for any significant dataset. Mongo will suck up drive space, load up a 140GB DB and you can end up using 200+ GB as the swap file grows during use.

And you're still going to want a fast drive.

In fact, I think it's safe to say the MongoDB is really a DB system that caters to leading-edge hardware (64-bit, lots of RAM, SSDs). I mean, the whole DB is centered around looking up data index data in RAM (hello 64-bit) and then doing focused random lookups on the drive (hello SSD).

why ... doesn't the entire industry jump ship from MySQL?

1. It's not ACID-compliant. Probably quite bad for the banking system (of course, most of them are still processing flat files, but that's a different issue). However, note that you can force "safe" writes with Mongo and guarantee that data gets to disk, but only one "document" at a time.
2. It's still very young. Lots of big business are still running old versions of Crystal Reports on their SQL Server 2000 app written in VB6. Or they're building enterprise service buses to manage the crazy heterogeneous environments they've built up over the years.
3. It's a very different paradigm. Maybe 30% of the questions I regularly see on Mongo mailing lists (and here) are fundamentally tied to "how do I do query X?" or "how do I structure this data?". Using MongoDB typically requires that you denormalize in advance. This is not only a little difficult, it's untrained. Most people only learn "normalization" in school, nobody teaches us how to denormalize for performance.
4. It's not the right tool for everything. Honestly I think that MongoDB is great tool for reading and writing transactional data. That simple "one-a-time" CRUD that comprises much of modern apps. However, MongoDB is not really great at reporting. In fact, I honestly envision that the next step is not "Mongo for everything" it's "Mongo for transactional" and "MySQL for reporting". When your data gets big enough that you throw out "real-time reporting", then using Map-Reduce to populate a reporting DB doesn't seem that bad.

As I understood it, as you scale, you get MySQL to feed Memcache. Now it appears I can start with something equally performant from the beginning.

Honestly, I'm working towards this on a few of my projects. Again, I think that MongoDB actually does make a valid caching layer. In fact, it makes a file-backed caching layer. So if you're capable of pushing MySQL change to Mongo, then you're getting getting Memcached without cache misses. It also makes it easy to "warm the cache" on new server, just copy files and start Mongo pointing at the correct folder, it really is that easy.