To do this he has to be using dynamic columns, and those are stored as one serialized blob per row. So the more data you have in the row, the more expensive the deserialization/reserialization is with each column you add. For very large series this could be an issue.
But it sounds like this is tolerable for his app because the writes are distributed over time in a predictable fashion.
I am a little surprised though at the author's claim that fetching a single big row results in "huge IO efficiency" over a range of small rows. I'd expect a small amount of overhead, but isn't it more or less the same amount of data being retrieved? What am I missing?
EDIT: I see the author mentioned that it reduces disk seeks because it's all serialized together already. Sort of like you're defragging the series data on every write. I guess that makes sense.
Personally I would probably look at using SSDs and keep the schema more "sane" and have more scalable writes, but that's just me.
In short, there is no deserialization/reserializaion. OP's writes are append only. I have a similar use pattern to OP, and I haven't seen any performance issues with 100,000s of columns (on SSDs)
2.) The "huge IO efficiency" is similar to what you would see in any columnar data store. Wikipedia has a good walkthrough of it (http://en.wikipedia.org/wiki/Column-oriented_DBMS). The short story is now there is fewer meta data between his values.
--
In any case, it works out because Cassandra is far more well suited for this type use pattern than Mongo is. We migrated from MongoDB (on SSDs) to Cassandra for similar reasons. The perf-killer on Mongo in this scenario is the write lock.
You are right. There is one row per symbol. However with Cassandra any given row can have any number of columns, so when you want to write a new value, you just create a new column for that second (/period).
The writes are not buffered in memory.