Cassandra Wide Vs Skinny Rows for large columns - performance

Question

I need to insert 60GB of data into cassandra per day.
This breaks down into
100 sets of keys
150,000 keys per set
4KB of data per key
In terms of write performance am I better off using
1 row per set with 150,000 keys per row
10 rows per set with 15,000 keys per row
100 rows per set with 1,500 keys per row
1000 rows per set with 150 keys per row
Another variable to consider, my data expires after 24 hours so I am using TTL=86400 to automate expiration
More specific details about my configuration:
CREATE TABLE stuff (
stuff_id text,
stuff_column text,
value blob,
PRIMARY KEY (stuff_id, stuff_column)
) WITH COMPACT STORAGE AND
bloom_filter_fp_chance=0.100000 AND
caching='KEYS_ONLY' AND
comment='' AND
dclocal_read_repair_chance=0.000000 AND
gc_grace_seconds=39600 AND
read_repair_chance=0.100000 AND
replicate_on_write='true' AND
populate_io_cache_on_flush='false' AND
compaction={'tombstone_compaction_interval': '43200', 'class': 'LeveledCompactionStrategy'} AND
compression={'sstable_compression': 'SnappyCompressor'};
Access pattern details:
The 4KB value is a set of 1000 4 byte floats packed into a string.
A typical request is going to need a random selection of 20 - 60 of those floats.
Initially, those floats are all stored in the same logical row and column. A logical row here represents a set of data at a given time if it were all written to one row with 150,000 columns.
As time passes some of the data is updated, within a logical row within the set of columns, a random set of levels within the packed string will be updated. Instead of updating in place, the new levels are written to a new logical row combined with other new data to avoid rewriting all of the data which is still valid. This leads to fragmentation as multiple rows now need to be accessed to retrieve that set of 20 - 60 values. A request will now typically read from the same column across 1 - 5 different rows.
Test Method
I wrote 5 samples of random data for each configuration and averaged the results. Rates were calculated as (Bytes_written / (time * 10^6)). Time was measured in seconds with millisecond precision. Pycassa was used as the Cassandra interface. The Pycassa batch insert operator was used. Each insert inserts multiple columns to a single row, insert sizes are limited to 12 MB. The queue is flushed at 12MB or less. Sizes do not account for row and column overhead, just data. The data source and data sink are on the same network on different systems.
Write results
Keep in mind there are a number of other variables in play due to the complexity of the Cassandra configuration.
1 row 150,000 keys per row: 14 MBps
10 rows 15,000 keys per row: 15 MBps
100 rows 1,500 keys per row: 18 MBps
1000 rows 150 keys per row: 11 MBps

Answer 1

The answer depends on what your data retrieval pattern is, and how your data is logically grouped. Broadly, here is what I think:
Wide row (1 row per set): This could be the best solution as it prevents the request from hitting several nodes at once, and with secondary indexing or composite column names, you can quickly filter data to your needs. This is best if you need to access one set of data per request. However, doing too many multigets on wide rows can increase memory pressure on nodes, and degrade performance.
Skinny row (1000 rows per set): On the other hand, a wide row can give rise to read hotspots in the cluster. This is especially true if you need to make a high volume of requests for a subset of data that exists entirely in one wide row. In such a case, a skinny row will distribute your requests more uniformly throughout the cluster, and avoid hotspots. Also, in my experience, "skinnier" rows tend to behave better with multigets.
I would suggest, analyze your data access pattern, and finalize your data model based on that, rather than the other way around.

Answer 2

You'd be better off using 1 row per set with 150,000 columns per row. Using TTL is good idea to have an auto-cleaning process.