Profiling looks at a data set and infers characteristics and constraints about the data.
- it helps users understand their data,
- inferred constraints may allow additional optimizations (e.g. a foreign key allows semi-join removal),
- column statistics help the optimizer estimate the selectivity of filters and joins,
- joint cardinalities drive the algorithm that chooses which tiles of a lattice to materialize.
Imagine you ran a profiler on a data set of 1 million rows and columns [orderId, gender, state, zipcode, productId, productName, brand]. Here is some sample output:
- the cardinality of 0 columns is the count of the data set;
- "nullCount" and "values" are only present for distributions of 1 column;
- "nullCount" may be is omitted if 0;
- "values" is only present if there are N or fewer values
- "distribution" of 2 or more columns is only output if it is "interesting"; in the case of ["state", "zipcode"] it is interesting because the joint cardinality 43,419 is fewer than the cardinality 999,982 that would be expected if they were uniform and independent;
- "fd" and "unique" are minimal. For example, we don't output "unique(orderId, productId)" if we have "unique(orderId)".
- Some measure of skewedness. Does one value occur many more times than others?
- Don't compute joint distributions for the power set of columns. This requires memory exponential in the number of columns. In pass 1 compute single-column distributions. In pass N compute N-column distributions only for the combinations of columns that the previous pass indicates will be interesting.
- Use HyperLogLog to compute cardinalities.
- Add low-cardinality columns to joint distributions. Rather than computing cardinality(zipcode, state) compute cardinality(zipcode, state, gender="F") and cardinality(zipcode, state, gender="M"). Because HLL rolls up losslessly, with 2x the memory you can compute 3 results: cardinality(zipcode, state), cardinality(zipcode, gender), cardinality(state, gender).
- Approximate histograms: approximate quartiles? Or buckets with exact counts in each range?
- Allow passing previous results into the algorithm. If you know the previous histogram of the orderDate column it is easier to compute its new histogram than if you start from scratch.
- HyperLogLog is inaccurate for low cardinalities, so keep all values until the number of values exceeds a threshold, then transition to buckets.