Demonstrates an approach to understanding performance cost of a single search
- 16G RAM, 8G might work when reducing JVM memory to 2G in docker-compose.yml
- 15G disk space to index in three different ways, SSD preferred
- Tested on Mac OS X, should work (better) on Linux
- Docker
- Python 3 with requests package
- Siege (available on most package managers) https://www.joedog.org/siege-manual/
- Download https://research.signalmedia.co/newsir16/signal-dataset.html
docker-compose up
You can now access Elasticsearch on http://localhost:9200 and Kibana on http://localhost:5601
Create indices and templates:
cd index-creation/
./create-articles-1-index.sh
./create-articles-5-index.sh
./create-articles-daily-template.sh
cd ..
Indexing:
- Indexing articles-1 and articles-5 took about 15min each on my Macbook
- Indexing articles-daily took 55 min
- The line progress counter will go up to 2 million
cd indexing/
./index-articles-1.sh
./index-articles-5.sh
./index-daily.sh
cd ..
## Do some testing
See what the example searches return:
./search.sh articles-1 vw-query-4.json
./search.sh articles-1 vw-query-16.json
For cleaner results, stop Kibana:
docker-compose stop kibana
This how to run all combinations of indices and searches. Do a warm-up run first, we don't care about IO.
./siege1000.sh articles-1 vw-query-4.json
./siege1000.sh articles-1 vw-query-16.json
./siege1000.sh articles-5 vw-query-4.json
./siege1000.sh articles-5 vw-query-16.json
./siege1000.sh articles-daily vw-query-4.json
./siege1000.sh articles-daily vw-query-16.json
The siege script disables the request cache. Check the indices.query_cache.hit_count metric and consider disabling the query cache for another round of testing:
./toggle-query-cache.sh articles-1 false
./toggle-query-cache.sh articles-5 false