This reminds me a bit of the self driving debate. AI might be safer than humans but because it's a robot we still prefer the human approach, even if overall it's more dangerous.
I do think there's something to the human element and that no one wants to take the fall but as companies find the value they will invest more in data quality and I suspect more and more will be handled upstream which should simplify the NL2SQL world. I think the semantic layer still needs to exist but I'm not sure where in the stack it will end up living.
I suppose I wasn't really thinking about if the data was actually wrong, more that it is perceived to be wrong. Definitely on pushing data quality upstream, which I wrote about a few weeks ago "We don't need Data Contracts". I think the semantic layer is it's own piece of the stack, it's where it is worked on that is the question... again I think the earlier and further upstream the better. Product engineers could actually define a good chunk of the semantic layer when doing normal product release work.
Ah got it. At some point it doesn't matter since perception is reality but when the data is knowingly bad it's a lot easier to dismiss the end result.
Interesting point. I do think the companies that do it well are putting data/analytics folks further up into the product development so they can influence the data that's collected and how it's collected but more often you see data on the receiving end of data and having to clean it up. But you still have multiple definitions since you will always have different consumers - finance will care about one thing, product teams another, sales something else, etc. In that world you should spend more time thinking about the data that's produced but still realize it's going to be transformed in a variety of ways for different use cases.
Yeah I remember when I was at Ocado (online grocery), marketing defined goods sold as anything ordered on the website, supply chain as anything that left the warehouse, trading as anything that left the warehouse and wasn’t returned and finance had the same less chargebacks for fraud! However I think these are all filters on the same metric, the requirement to be able to differentiate should be in the product spec.
This reminds me a bit of the self driving debate. AI might be safer than humans but because it's a robot we still prefer the human approach, even if overall it's more dangerous.
I do think there's something to the human element and that no one wants to take the fall but as companies find the value they will invest more in data quality and I suspect more and more will be handled upstream which should simplify the NL2SQL world. I think the semantic layer still needs to exist but I'm not sure where in the stack it will end up living.
I suppose I wasn't really thinking about if the data was actually wrong, more that it is perceived to be wrong. Definitely on pushing data quality upstream, which I wrote about a few weeks ago "We don't need Data Contracts". I think the semantic layer is it's own piece of the stack, it's where it is worked on that is the question... again I think the earlier and further upstream the better. Product engineers could actually define a good chunk of the semantic layer when doing normal product release work.
Ah got it. At some point it doesn't matter since perception is reality but when the data is knowingly bad it's a lot easier to dismiss the end result.
Interesting point. I do think the companies that do it well are putting data/analytics folks further up into the product development so they can influence the data that's collected and how it's collected but more often you see data on the receiving end of data and having to clean it up. But you still have multiple definitions since you will always have different consumers - finance will care about one thing, product teams another, sales something else, etc. In that world you should spend more time thinking about the data that's produced but still realize it's going to be transformed in a variety of ways for different use cases.
Yeah I remember when I was at Ocado (online grocery), marketing defined goods sold as anything ordered on the website, supply chain as anything that left the warehouse, trading as anything that left the warehouse and wasn’t returned and finance had the same less chargebacks for fraud! However I think these are all filters on the same metric, the requirement to be able to differentiate should be in the product spec.