Cache recommendations in Data Virtualization
Using an input set of queries, Data Virtualization recommends a ranked list of data caches that can improve the performance of the input queries and potentially help future query workloads.
The input queries are queries that were run anywhere in the previous 1 day up to the previous 15 days, and must have an execution time of at least one minute. The recommendations are considered valid for 1 day after which they can change as the query workload changes.
- The cache recommendation engine uses two models to generate recommendations.
- The rule-based model uses sophisticated heuristics to determine which cache candidates help the input query workload.
- The machine learning based model uses a pre-trained machine learning model that detects underlying query patterns and predicts caches that help a potential future query workload.
Both models produce a ranked list of cache candidates that are consolidated by the engine to generate a final set of recommendations. You can choose to enable or disable machine learning based cache recommendations. By default, machine learning based cache recommendations are enabled.
In addition to cache creation recommendations, the engine also recommends cache disable and delete recommendations based on past usage and other metrics. These recommendations appear in the Active and Inactive caches tab for existing caches.
- Machine learning based cache recommendations consider underlying query patterns and predict
caches that are valid for 1 day.
Data Virtualization uses a pre-trained model that was trained on an industry standard data set.
You can choose to enable or disable machine learning based cache recommendations.
- The recommendation engine consolidates and ranks the final set of recommendations from both models. The Manager can then add data caches from these recommendations.
Data Virtualization provides an engine to generate a ranked list of recommendations. Ranking of cache creation recommendations is determined by the execution time of queries, the frequency of those queries in the input workload, and the weight of the two models. The engine is fully aware and does not recommend the creation of caches that exist. Additionally, the engine does not recommend the creation of duplicated caches.
The process of generating cache recommendations consists of five stages, as depicted in the following image:
- Collect
- The cache recommendation engine collects information such as query text, execution time,
cardinality, timestamp, and frequency for the provided time period.The following image shows how queries from historical workload are filtered to arrive to the final input set of queries for the recommendation engine: Cache configuration settings are defined by the Data Virtualization Manager, see Configure cache recommendations for details.
- Extract
- The recommendation engine generates potential cache candidates for the input query workload.
- Translate
- The recommendation engine converts and consolidates the candidates to ensure that they are syntactically and semantically correct, unique, and pass all Db2® restrictions.
- Evaluate
- The engine evaluates the converted definitions by matching each cache candidate against the
input query workload. Also, for scoring the machine learning model, a high-dimensional feature
vector is created for each candidate.
As a result of this evaluation, the recommendation engine generates a match score for each cache candidate. The evaluation of each candidate is based on the following criteria.
- Matchability Number of queries that match the cache candidate.
- Diversity Different queries that match the cache candidate.
- Cardinality Size of result set that the cache candidate fetched.
- Performance Execution time of queries that the cache candidate matched.
- Rank and sort
- The recommendation engine ranks and sorts the cache candidates to generate a final list of
recommendations. The final list of recommendations is created based on the following criteria.
- Sort candidates by using a weighted metric of matched query execution time and frequency.
- Any tie between candidates is broken by using query frequency and cardinality.