You can create a temporary table and write some common JSONB documents to it manually or use existing tables. The idea is to provide a subset of real data. Lets say some document*type* is twice as frequent as some other document type. ZSON expects that there will be twicemoredocuments of the first typethanof the second in a learning set.

You can create a temporary table and write some common JSONB documents to it manually or use existing tables. The idea is to provide a subset of real data. Lets say some document*type* is twice as frequent as some other document type. ZSON expects that there will be twiceas manydocuments of the first typeas thoseof the second one in a learning set.

Resulting dictionary could be examined using this query:

When schema of JSONB documents evolve ZSON could be*re-learned*:

This time*second* dictionary will be created. Dictionaries are cached in memory so it will take about a minute before ZSON realizes that there is a new dictionary. After that old documents will be decompressed using old dictionary and new documents will be compressed and decompressed using new dictionary.

This time*second* dictionary will be created. Dictionaries are cached in memory so it will take about a minute before ZSON realizes that there is a new dictionary. After that old documents will be decompressed usingtheold dictionary and new documents will be compressed and decompressed using the new dictionary.

To find out which dictionary is used for given ZSON document use zson_info procedure:

To find out which dictionary is used foragiven ZSON document use zson_info procedure:

If**all** ZSON documents are migrated to new dictionary the old one could be safely removed:

If**all** ZSON documents are migrated tothenew dictionary the old one could be safely removed:

In this case ZSON gives about 11.8% more TPS.

We can modify compress.pgbench and nocompress.pgbench so only documents with idinbetween of 1 and 3000 will be requested. It will simulate a case when all data*does* fit into memory. In this case we see 141K TPS (JSONB) vs 134K TPS (ZSON) which is 5% slower.

We can modify compress.pgbench and nocompress.pgbench so onlythedocuments with id between 1 and 3000 will be requested. It will simulate a case when all data*does* fit into memory. In this case we see 141K TPS (JSONB) vs 134K TPS (ZSON) which is 5% slower.

Compressionratio could be different depending on documents, database schema, number of rows, etc. But in general ZSON compression is much better than build-in PostgreSQL compression (PGLZ):

The compressionratio could be different depending onthedocuments,thedatabase schema, the number of rows, etc. But in general ZSON compression is much better than build-in PostgreSQL compression (PGLZ):