Brief comments on the purpose of your changes:
For Dune Engine V2
I've checked that:
- I tested the query on dune.com after compiling the model with dbt compile (compiled queries are written to the target directory)
- I used "refs" to reference other models in this repo and "sources" to reference raw or decoded tables
- if adding a new model, I added a test
- the filename is unique and ends with .sql
- each sql file is a select statement and has only one view, table or function defined
- column names are
lowercase_snake_cased
- if adding a new model, I edited the dbt project YAML file with new directory path for both models and seeds (if applicable)
- if wanting to expose a model in the UI (Dune data explorer), I added a post-hook in the JINJA config to add metadata (blockchains, sector/project, name and contributor Dune usernames)
-
coin_id
represents the ID of the coin on coinpaprika.com - all the coins are active on coinpaprika.com (please remove inactive ones)
- if joining to base table (i.e. ethereum transactions or traces), I looked to make it an inner join if possible
- I used is_incremental & not is_incremental jinja block filters on both base tables and decoded tables
- where block_time >= date_trunc("day", now() - interval '1 week')
- if joining to base table (i.e. ethereum transactions or traces), I applied join condition where block_time >= date_trunc("day", now() - interval '1 week')
- if joining to prices view, I applied join condition where minute >= date_trunc("day", now() - interval '1 week')