5 ms·

A decent cost based optimizer will be right most of the time, and you have hints for the cases where it falters. Your approach is simply hiding the query optim

by sergei 13y ago

A decent cost based optimizer will be right most of the time, and you have hints for the cases where it falters. Your approach is simply hiding the query optimizer choices inside of the app, making it more fragile. Let's say I have two fields: a and b, both of which have an index. And then I have a query with predicates across both a and b. For example, a = 10 and b = 20. One of the core facilities within an optimizer is selectivity estimation. By looking at the statistics, the optimizer will see that a = 10 might look at 10,000 rows while b = 20 might only look at 10. So the optimal and desired choice would be to use index over b. However, the exact same query construct with different parameters (e.g. a = 50 and b = 3) might flip the index selection. Now let's imagine I had to implement this inside of my app. Every time I have such a constraint, the app has to become aware of selectivity to know which indexes to use based on input parameters.

josegonzalez 13y ago
Which is good, because now your developers don't just add queries to the application without thinking about performance. You don't have to be a dba to understand how using one index vs another will affect the performance - and thus conversion rate - of your application.
- dennis82 13y ago
  no, it's categorically not good to make an application more fragile. A weakness restated is not a strength. Every database should have an optimizer, period.
  - josegonzalez 13y ago
    I did not state that an optimizer should not exist for a database - I think thats key actually - but rather that the tradeoff they made this time around was fundamentally good in that - at least for now - it forces the developer to think about application performance. If that happens to make an application more fragile, I think that is more of a code organization/tooling issue than anything else.
  - pkolaczk 13y ago
    If a database system does not support joins, aggregations or subqueries like most realtime NoSQL solutions do, an optimizer becomes pretty trivial. Optimizers are needed for analytical stuff. That's why most optimizers are evaluated on analytic workloads (e.g. TPC-H, TPC-W) not transactional / realtime (TPC-C).
- sergei 13y ago
  In practice, you end up with apps which dynamically build up predicates in different sections of the code. And when you combine that with many predicates, many tables, and other constraints such as ordering or aggregates, things get complex pretty quick. Even if you have the best developers who understand all the in an outs of the dataset, re-implementing an optimizer in the app is rarely the right choice.
coffeemug 13y ago
slava @ rethink here. I don't think we disagree -- optimizers are unquestionably a good idea and are immensely useful. For us it was a matter of a) giving people the option to specify indexes directly, and b) shipping quickly. We'll implement a proper statistical optimizer in due time, but for the time being we found that the explicit approach gives a lot of people 80% of what they need with 20% of the work on our part.
btilly 13y ago
You are underestimating the necessary design tension. If you expose hints, then you've constrained how your optimizer works because it has to work with those hints. (Oracle, for example, has been fighting this battle for ages.) Once you've built an optimizer, people come to depend on it so it is critical that you get it right. They will need optimizer eventually, and they know it. But I'm glad that they are not implementing it before they are ready. Secondly your selectivity estimation point can go either way, and on the whole I don't like it. I've personally experienced the situation where a database recomputes statistics, the CBO decides that a new query plan is needed for a common query, it made a poor choice, and the first that any human hears about it, the site is down. The problem here is that the risk profile for the application of trying to be smart here is completely backwards. In general, as long as a query is not a bottleneck, I don't care about making it faster. Oh, you made it 2x faster? I didn't mind before and I'm unlikely to even know that you did so. But if you JUST ONCE switch to a bad plan on production without warning, your users WILL notice, they WILL care, and they WILL NOT be happy. As a developer, I don't care that you make the right choice 95% of the time. I want you to make a predictable choice. Because if you're making a bad choice in development, I've got a window of opportunity to notice and do something about it where nobody cares. But if you randomly make changes on production, every single mistake counts. Oh, but you say that this just means that you need stored query plans? I agree, and this is an example of why the behavior of the optimizer has to be thought through very carefully before you just throw something out there, people come to depend on it, and then you realize that you have put barriers to thinking of it the way you want to think of it.
- sergei 13y ago
  I frequently see apps with thousands of distinct query signatures. Having a developer manually chose indexes, join ordering, aggregation method, (just to name a few) for every single query, and then select multiple plans because input parameters absolutely do result in scenario where plans can be 10,000x off in performance -- well, that's just untenable. Not trying to claim that CBOs are the panacea here, but let's be realistic. Having developers manually plan every single query is not the right choice.
  - 13y ago
jdoliner 13y ago
An optimizer can be right most of the time but it's basically impossible to have it be right all of the time. It really depends on what you're doing as to which approach you prefer. If you absolutely need a query to have a certain performance characteristic then you can't count on the optimizer to get it right a human needs to think through exactly how they want this query to run to ensure it behaves correctly. There definitely are several use cases where people are willing to sacrifice a risk of misoptimization in exchange for shorter code and we fully intend to support that (it's a complicated project though so we don't now.) However I disagree that having this logic in the app makes it more fragile. Having the precise semantics expressed in the application means you always know exactly how a query is going to be performed. We know this execution plan isn't going to change based on seemingly unrelated properties of the data (as it does with optimizers.) If fragility is the propensity to break then I'd argue having the optimizer in control rather than the developer gives you more fragility than less.

by sergei 13y ago

josegonzalez 13y ago

Which is good, because now your developers don't just add queries to the application without thinking about performance. You don't have to be a dba to understand how using one index vs another will affect the performance - and thus conversion rate - of your application.

dennis82 13y ago
no, it's categorically not good to make an application more fragile. A weakness restated is not a strength. Every database should have an optimizer, period.
- josegonzalez 13y ago
  I did not state that an optimizer should not exist for a database - I think thats key actually - but rather that the tradeoff they made this time around was fundamentally good in that - at least for now - it forces the developer to think about application performance. If that happens to make an application more fragile, I think that is more of a code organization/tooling issue than anything else.
- pkolaczk 13y ago
  If a database system does not support joins, aggregations or subqueries like most realtime NoSQL solutions do, an optimizer becomes pretty trivial. Optimizers are needed for analytical stuff. That's why most optimizers are evaluated on analytic workloads (e.g. TPC-H, TPC-W) not transactional / realtime (TPC-C).
sergei 13y ago
In practice, you end up with apps which dynamically build up predicates in different sections of the code. And when you combine that with many predicates, many tables, and other constraints such as ordering or aggregates, things get complex pretty quick. Even if you have the best developers who understand all the in an outs of the dataset, re-implementing an optimizer in the app is rarely the right choice.

coffeemug 13y ago

slava @ rethink here. I don't think we disagree -- optimizers are unquestionably a good idea and are immensely useful. For us it was a matter of a) giving people the option to specify indexes directly, and b) shipping quickly. We'll implement a proper statistical optimizer in due time, but for the time being we found that the explicit approach gives a lot of people 80% of what they need with 20% of the work on our part.

btilly 13y ago

You are underestimating the necessary design tension. If you expose hints, then you've constrained how your optimizer works because it has to work with those hints. (Oracle, for example, has been fighting this battle for ages.) Once you've built an optimizer, people come to depend on it so it is critical that you get it right. They will need optimizer eventually, and they know it. But I'm glad that they are not implementing it before they are ready. Secondly your selectivity estimation point can go either way, and on the whole I don't like it. I've personally experienced the situation where a database recomputes statistics, the CBO decides that a new query plan is needed for a common query, it made a poor choice, and the first that any human hears about it, the site is down. The problem here is that the risk profile for the application of trying to be smart here is completely backwards. In general, as long as a query is not a bottleneck, I don't care about making it faster. Oh, you made it 2x faster? I didn't mind before and I'm unlikely to even know that you did so. But if you JUST ONCE switch to a bad plan on production without warning, your users WILL notice, they WILL care, and they WILL NOT be happy. As a developer, I don't care that you make the right choice 95% of the time. I want you to make a predictable choice. Because if you're making a bad choice in development, I've got a window of opportunity to notice and do something about it where nobody cares. But if you randomly make changes on production, every single mistake counts. Oh, but you say that this just means that you need stored query plans? I agree, and this is an example of why the behavior of the optimizer has to be thought through very carefully before you just throw something out there, people come to depend on it, and then you realize that you have put barriers to thinking of it the way you want to think of it.

sergei 13y ago
I frequently see apps with thousands of distinct query signatures. Having a developer manually chose indexes, join ordering, aggregation method, (just to name a few) for every single query, and then select multiple plans because input parameters absolutely do result in scenario where plans can be 10,000x off in performance -- well, that's just untenable. Not trying to claim that CBOs are the panacea here, but let's be realistic. Having developers manually plan every single query is not the right choice.
- 13y ago

jdoliner 13y ago

An optimizer can be right most of the time but it's basically impossible to have it be right all of the time. It really depends on what you're doing as to which approach you prefer. If you absolutely need a query to have a certain performance characteristic then you can't count on the optimizer to get it right a human needs to think through exactly how they want this query to run to ensure it behaves correctly. There definitely are several use cases where people are willing to sacrifice a risk of misoptimization in exchange for shorter code and we fully intend to support that (it's a complicated project though so we don't now.) However I disagree that having this logic in the app makes it more fragile. Having the precise semantics expressed in the application means you always know exactly how a query is going to be performed. We know this execution plan isn't going to change based on seemingly unrelated properties of the data (as it does with optimizers.) If fragility is the propensity to break then I'd argue having the optimizer in control rather than the developer gives you more fragility than less.