[
https://issues.apache.org/jira/browse/PIG-143?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12577649#action_12577649
]
Alan Gates commented on PIG-143:
--------------------------------
Pi,
I am totally in agreement with your suggestions. I agree that performance is
not a huge issue as most plans will be represented by graphs of 1000 nodes or
less.
In the types branch we have already made some progress in this direction. I
suspect therefore it would be easiest for you to implement these changes (if
you are interested in doing so) in the types branch. We're reworking the
logical plan to be a generic graph and have implemented visitation methods that
traverse the graph depth first and in dependency order. My previous experience
has been that these two together (depth first for speed, dependency order when
you need to propagate state (such as in a type checker)) are sufficient for the
needs of all types of visitors. I haven't checked in the changes to the
logical plan yet, but under pig/impl/plan/OperatorPlan.java and associated
classes you can see the generic graph package we wrote. The logical plan will
extend this. I hope to have changes to the logical plan that match this
checked in within a couple of days. Take a look and let me know what you think.
> Proposal for refactoring of parsing logic in Pig
> ------------------------------------------------
>
> Key: PIG-143
> URL: https://issues.apache.org/jira/browse/PIG-143
> Project: Pig
> Issue Type: Improvement
> Components: impl
> Reporter: Pi Song
> Assignee: Pi Song
> Attachments: ParserDrawing.png
>
>
> h2. Pig Script Parser Refactor Proposal
> This is my initial proposal on pig script parser refactor work. Please note
> that I need your opinions for improvements.
> *Problem*
> The basic concept is around the fact that currently we do validation logics
> in parsing stage (for example, file existence checking) which I think is not
> clean and difficult to add new validation rules. In the future, we will need
> to add more and more validation logics to improve usability.
> *My proposal:-* (see [^ParserDrawing.png])
> - Only keep parsing logic in the parser and leave output of parsing logic
> being unchecked logical plans. (Therefore the parser only does syntactic
> checking)
> - Create a new class called LogicalPlanValidationManager which is responsible
> for validations of the AST from the parser.
> - A new validation logic will be subclassing LogicalPlanValidator
> - We can chain a set of LogicalPlanValidators inside
> LogicalPlanValidationManager to do validation work. This allows a new
> LogicalPlanValidator to be added easily like a plug-in.
> - This greatly promotes modularity of the validation logics which is
> +particularly good when we have a lot of people working on different things+
> (eg. streaming may require a special validation logic)
> - We can set the execution order of validators
> - There might be some backend specific validations needed when we implement
> new execution engines (For example a logical operation that one backend can
> do but others can't). We can plug-in this kind of validations on-the-fly
> based on the backend in use.
> *List of LogicalPlanValidators extracted from the current parser logic:-*
> - File existence validator
> - Alias existence validator
> *Logics possibly be added in the very near future:-*
> - Streaming script test execution
> - Type checking + casting promotion + type inference
> - Untyped plan test execution
> - Logic to prevent reading and writing from/to the same file
> The common way to implement a LogicalPlanValidator will be based on Visitor
> pattern.
> *Cons:-*
> - By having every validation logic traversing AST from the root node every
> time, there is a performance hit. However I think this is neglectable due to
> the fact that Pig is very expressive and normally queries aren't too big (99%
> of queries contain no more than 1000 AST nodes).
> *Next Step:-*
> LogicalPlanFinalizer which is also a pipeline except that each stage can
> modify the input AST. This component will generally do a kind of global
> optimizations.
> *Further ideas:-*
> - Composite visitor can make validations more efficient in some cases but I
> don't think we need
> - ASTs within the pipeline never change (read-only) so validations can be
> done in parallel to improve responsiveness. But again I don't think we need
> this unless we have so many I/O bound logics.
> - The same pipeline concept can also be applied in physical plan
> validation/optimization.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
