Some time ago I thought an new statistics system over, for our multi-million user website, to log and report user-actions for our customers.
The database-design is quite simple, containing one table, with a foreignId (200,000 different id’s), a datetime field, an actionId (30 different id’s), and two more fields containing some meta-information (just smallints). There are no constraints to other tables. Furthermore we have two indexes each containing 4 fields, which cannot be dropped, as users are getting timeouts when we are having smaller indexes. The foreignId is the most important field, as each and every query contains this field.
We chose to use SQL server, but after implementation doesn’t a relational database seem like a perfect fit, as we cannot insert 30 million records a day (it’s insert only, we don’t do any updates) when also doing alot of random reads on the database; because the indexes cannot be updated fast enough. Ergo: we have a massive problem 🙂 We have temporarily solved the problem, yet
a relational database doesn’t seem to be suited for this problem!
Would a database like BigTable be a better choice, and why? Or are there other, better choices when dealing with this kind of problems?
NB. At this point we use a single 8-core Xeon system with 4 GB memory and Win 2003 32-bit. RAID10 SCSI as far as I know. The index size is about 1.5x the table size.
You say that your system is capable of inserting 3000 records per second without indexes, but only about 100 with two additional non-clustered indexes. If 3k/s is the maximum throughput your I/O permits, adding two indexes should in theory reduces the throughput at about 1000-1500/sec. Instead you see a degradation 10 times worse. The proper solution and answer is ‘It Dependts’ and some serious troubleshooting and bottleneck identification would have to be carried out. With that in mind, if I was to venture a guess, I’d give two possible culprits:
A. Th additional non-clustered indexes distribute the writes of dirty pages into more allocation areas. The solution would be to place the the clustered index and each non-clustered index into its own filegroup and place the three filegroups each onto separate LUNs on the RAID.
B. The low selectivity of the non-clustered indexes creates high contention between reads and writes (key conflicts as well as %lockres% conflicts) resulting in long lock wait times for both inserts and selects. Possible solutions would be using SNAPSHOTs with read committed snapshot mode, but I must warn about the danger of adding lot of IO in the version store (ie. in tempdb) on system that may already be under high IO stress. A second solution is using database snapshots for reporting, they cause lower IO stress and they can be better controlled (no tempdb version store involved), but the reporting is no longer on real-time data.
I tend to believe B) as the likely cause, but I must again stress the need to proper investigation and proper root case analysis.
‘RAID10’ is not a very precise description.
As on the question whether relational databases are appropriate for something like this, yes, absolutely. There are many more factors to consider, recoverability, availability, toolset ecosystem, know-how expertise, ease of development, ease of deployment, ease of management and so on and so forth. Relational databases can easily handle your workload, they just need the proper tuning. 30 million inserts a day, 350 per second, is small change for a database server. But a 32bit 4GB RAM system hardly a database server, regardless the number of CPUs.