[
https://issues.apache.org/jira/browse/TINKERPOP-2753?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Boxuan Li updated TINKERPOP-2753:
---------------------------------
Description:
I only have experience in JanusGraph, so my opinion might be biased and this
proposal might not be generalizable to other graph providers:
I propose we create a `noop()` step that does nothing. It is a special step
that simply provides a hint for the graph provider. How to interpret it depends
on the graph provider, but the usage in my mind is to avoid eager optimization.
Sometimes a graph provider can combine different filter steps into a joint
condition for better index selection or predicate pushdown. For example, in the
query below:
```
g.V().has("name", "bob").has("age", 20)
```
JanusGraph will fold the two `has` conditions into a joint condition for better
index selection. Sometimes, however, users don't want this "eager
optimization", likely because they know the distribution of data and prefer
doing in-memory filtering for the second `has` condition. They could do this:
```java
g.V().has("name", "bob").map(x -> x.get()).has("age", 20)
```
So that JanusGraph will defer the evaluation of the second condition until the
first `has` condition is evaluated. Here, the `map(x -> x.get())` is
essentially a noop step. What I am proposing is to use an official `noop()`
step to replace this workaround. This `noop` step sounds like a `barrier` step
but they do not have the same semantics. The `noop` step is a barrier against
constraint look-ahead optimization.
Another example usage of `noop` is as follows:
```java
g.V(ids).bothE("follows").noop().where(__.otherV().is(v2)).next()
```
In the above case, we can use `noop` to force the graph provider to compute
`bothE` first and then evaluate `where` statement. Otherwise, the graph
provider (for example, JanusGraph) will try folding the `where` condition into
the `bothE` step for predicate pushdown. Predicate pushdown usually works, but
in some scenarios, it is less preferred.
I am happy to provide a patch if the community likes this idea.
was:
I only have experience in JanusGraph, so my opinion might be biased and this
proposal might not be generalizable to other graph providers:
I propose we create a `noop()` step that does nothing. It is a special step
that simply provides a hint for the graph provider. How to interpret it depends
on the graph provider, but the usage in my mind is to avoid eager optimization.
Sometimes a graph provider can combine different filter steps into a joint
condition for better index selection or predicate pushdown. For example, in the
query below:
```java
g.V().has("name", "bob").has("age", 20)
```
JanusGraph will fold the two `has` conditions into a joint condition for better
index selection. Sometimes, however, users don't want this "eager
optimization", likely because they know the distribution of data and prefer
doing in-memory filtering for the second `has` condition. They could do this:
```java
g.V().has("name", "bob").map(x -> x.get()).has("age", 20)
```
So that JanusGraph will defer the evaluation of the second condition until the
first `has` condition is evaluated. Here, the `map(x -> x.get())` is
essentially a noop step. What I am proposing is to use an official `noop()`
step to replace this workaround. This `noop` step sounds like a `barrier` step
but they do not have the same semantics. The `noop` step is a barrier against
constraint look-ahead optimization.
Another example usage of `noop` is as follows:
```java
g.V(ids).bothE("follows").noop().where(__.otherV().is(v2)).next()
```
In the above case, we can use `noop` to force the graph provider to compute
`bothE` first and then evaluate `where` statement. Otherwise, the graph
provider (for example, JanusGraph) will try folding the `where` condition into
the `bothE` step for predicate pushdown. Predicate pushdown usually works, but
in some scenarios, it is less preferred.
I am happy to provide a patch if the community likes this idea.
> Create noop() step to avoid eager optimization
> ----------------------------------------------
>
> Key: TINKERPOP-2753
> URL: https://issues.apache.org/jira/browse/TINKERPOP-2753
> Project: TinkerPop
> Issue Type: New Feature
> Reporter: Boxuan Li
> Priority: Major
>
> I only have experience in JanusGraph, so my opinion might be biased and this
> proposal might not be generalizable to other graph providers:
> I propose we create a `noop()` step that does nothing. It is a special step
> that simply provides a hint for the graph provider. How to interpret it
> depends on the graph provider, but the usage in my mind is to avoid eager
> optimization. Sometimes a graph provider can combine different filter steps
> into a joint condition for better index selection or predicate pushdown. For
> example, in the query below:
> ```
> g.V().has("name", "bob").has("age", 20)
> ```
> JanusGraph will fold the two `has` conditions into a joint condition for
> better index selection. Sometimes, however, users don't want this "eager
> optimization", likely because they know the distribution of data and prefer
> doing in-memory filtering for the second `has` condition. They could do this:
> ```java
> g.V().has("name", "bob").map(x -> x.get()).has("age", 20)
> ```
> So that JanusGraph will defer the evaluation of the second condition until
> the first `has` condition is evaluated. Here, the `map(x -> x.get())` is
> essentially a noop step. What I am proposing is to use an official `noop()`
> step to replace this workaround. This `noop` step sounds like a `barrier`
> step but they do not have the same semantics. The `noop` step is a barrier
> against constraint look-ahead optimization.
>
> Another example usage of `noop` is as follows:
> ```java
> g.V(ids).bothE("follows").noop().where(__.otherV().is(v2)).next()
> ```
> In the above case, we can use `noop` to force the graph provider to compute
> `bothE` first and then evaluate `where` statement. Otherwise, the graph
> provider (for example, JanusGraph) will try folding the `where` condition
> into the `bothE` step for predicate pushdown. Predicate pushdown usually
> works, but in some scenarios, it is less preferred.
>
> I am happy to provide a patch if the community likes this idea.
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
