[dietr1ch]

Analytics is done in "batches" (daily, weekly) anyways, right?

We know you can't get both, row and column orders at the same time, and that continuously maintaining both means duplication and ensuring you get the worst case from both worlds.

Local, row-wise writing is the way to go for write performance. Column-oriented reads are the way to do analytics at scale. It seems alright to have a sync process that does the order re-arrangement (maybe with extra precomputed statistics, and sharding to allow many workers if necessary) to let queries of now historical data run fast.

[delaminator]

It's not just about row versus column. OLAPs are potentially denormalised as well, and sometimes pre-aggregation, such as rolling up by day, by customer.

If you really need to get performance you'll be building a star schema.

[sanderjd]

Not all olap-like queries are for daily reporting.

I agree that the basic architecture should be row order -> delay -> column order, but the question (in my mind) is balancing the length of that delay with the usefulness of column order queries for a given workload. I seem to keep running into workloads that do inserts very quickly and then batch reads on a slower cadence (either in lockstep with the writes, or concurrently) but not on the extremely slow cadence seen in the typical olap reporting type flow. Essentially, building up state and then querying the results.

I'm not so sure about "continuously maintaining both means duplication and ensuring you get the worst case from both worlds". Maybe you're right, I'm just not so sure. I agree that it's duplicating storage requirements, but is that such a big deal? And I think if fast writes and lookups and fast batch reads are both possible at the cost of storage duplication, that would actually be the best case from both worlds?

I mean, this isn't that different conceptually from the architecture of log-structured merge trees, which have this same kind of "duplication" but for good purpose. (Indeed, rocksdb has been the closest thing to what I want for this workload that I've found; I just think it would be neat if I could use sqlite+duckdb instead, accepting some tradeoffs.)

[dietr1ch]

> the question (in my mind) is balancing the length of that delay with the usefulness of column order queries for a given workload. I seem to keep running into workloads that do inserts very quickly and then batch reads on a slower cadence (either in lockstep with the writes, or concurrently) but not on the extremely slow cadence seen in the typical olap reporting type flow. Essentially, building up state and then querying the results.

I see. Can you come up with row/table watermarks? Say your column store is up-to-date with certain watermark, so any query that requires freshness beyond that will need to snoop into the rows that haven't made it into the columnar store to check for data up to the required query timestamp.

In the past I've dealt with a system that had read-optimised columnar data that was overlaid with fresh write-optimised data and used timestamps to agree on the data that should be visible to the queries. It continuously consolidated data into the read-optimised store instead of having the silly daily job that you might have in the extremely slow cadence reporting job you mention.

You can write such a system, but in reality I've found it hard to justify building a system for continuous updates when a 15min delay isn't the end of the world, but it's doable if you want it.

> I'm not so sure about "continuously maintaining both means duplication and ensuring you get the worst case from both worlds". Maybe you're right, I'm just not so sure. I agree that it's duplicating storage requirements, but is that such a big deal? And I think if fast writes and lookups and fast batch reads are both possible at the cost of storage duplication, that would actually be the best case from both worlds?

I mean that if you want both views in a consistent world, then writes will bring things to a crawl as both, row and column ordered data needs to be updated before the writing lock is released.

[sanderjd]

Yes! We're definitely talking about the same thing here! Definitely not thinking of consistent writes to both views.

Now that you said this about watermarks, I realize that this is definitely the same idea as streaming systems like flink (which is where I'm familiar with watermarks from), but my use cases are smaller data and I'm looking for lower latency than distributed systems like that. I'm interested in delays that are on the order of double to triple digit milliseconds, rather than 15 minutes. (But also not microseconds.)

I definitely agree that it's difficult to justify building this, which is why I keep looking for a system that already exists :)