MAT Calcite plugin

About

This plugin for Eclipse Memory Analyzer allows to query heap dump via SQL

While MAT does have a query language, it does NOT allow to join, sort and group results. MAT Calcite plugin allows all the typical SQL operations (joins, filters, group by, order by, etc)

Query engine is implemented via Apache Calcite See Calcite SQL reference

Installation

Requirements: Java 17+, Eclipse Memory Analyzer 0.14.0+

There are two update channels:

• https://vlsi.github.io/mat-calcite-plugin-update-site/stable (stable releases)
• https://vlsi.github.io/mat-calcite-plugin-update-site/ea (early access, per-commit releases)

To install Calcite SQL plugin, perform the following:

1. Open Help, Install New Software...
2. Click Add, it will open a Add Repository window
3. Type Calcite SQL plugin stable releases to the Name field
4. Type https://vlsi.github.io/mat-calcite-plugin-update-site/stable/ to the Location field
5. Click Ok
6. All the checkboxes can be left by default (Show only latest version, Group items by category, ...)
7. Check SQL for Memory Analyzer category
8. Click Next (Available Software)
9. Click Next (Installation Details)
10. Accept License
11. Click Finish and restart MAT

Early access versions

Development builds are pushed to https://vlsi.github.io/mat-calcite-plugin-update-site/ea/ repository, so you can preview the upcoming version right after the commit lands to the default branch.

Sample

Query that lists duplicate URLs:

select toString(file) file_str, count(*) cnt, sum(retainedSize(this)) sum_retained, sum(shallowSize(this)) sum_shallow
  from java.net.URL
 group by toString(file)
having count(*)>1
 order by sum(retainedSize(this)) desc

To get an explain plan, use "explain plan for select ...":

EnumerableSort(sort0=[$2], dir0=[DESC])
  View (expr#0..3=[{inputs}], expr#4=[1], expr#5=[>($t1, $t4)], proj#0..3=[{exprs}], $condition=[$t5])
    EnumerableAggregate(group=[{0}], cnt=[COUNT()], sum_retained=[$SUM0($1)], sum_shallow=[$SUM0($2)])
      View (expr#0=[{inputs}], expr#1=[0], expr#2=[GET_SNAPSHOT($t1)], expr#3=[GET_IOBJECT($t2, $t0)], expr#4=['file'], expr#5=[RESOLVE_REFERENCE($t3, $t4)], expr#6=[toString($t5)], expr#7=[TO_REFERENCE($t3)], expr#8=[retainedSize($t7)], expr#9=[shallowSize($t7)], file=[$t6], $f1=[$t8], $f2=[$t9])
        GetObjectIdsByClass (class=java.net.URL)

Join sample

 select u.this, retainedSize(s.this)
   from "java.lang.String" s
   join "java.net.URL" u
     on s.this = u.path

Here's execution plan:

View (expr#0..2=[{inputs}], expr#3=[retainedSize($t2)], this=[$t0], EXPR$1=[$t3])
  HashJoin (condition=[=($1, $2)], joinType=[inner])
    View (expr#0=[{inputs}], expr#1=[0], expr#2=[GET_SNAPSHOT($t1)], expr#3=[GET_IOBJECT($t2, $t0)], expr#4=[TO_REFERENCE($t3)], expr#5=['path'], expr#6=[RESOLVE_REFERENCE($t3, $t5)], this=[$t4], path=[$t6])
      GetObjectIdsByClass (class=java.net.URL)
    View (expr#0=[{inputs}], expr#1=[0], expr#2=[GET_SNAPSHOT($t1)], expr#3=[GET_IOBJECT($t2, $t0)], expr#4=[TO_REFERENCE($t3)], this=[$t4])
      GetObjectIdsByClass (class=java.lang.String)

Heap schema

heap (default schema)
+- java (sub-schema name)
   +- util (sub-schema name)
      +- HashMap (table name).
         This "table" would return all the instances of java.util.HashMap without subclasses
+- instanceof
   +- java
      +- util
         +- HashMap (table name).
            This would return HashMap instances as well as subclass instances (e.g. LinkedHashMap)
+- "java.util.HashMap" (table name)
   This can be used as alternative.
+- native.ThreadStackFrames (table name)
   Returns thread stack traces and local variable info

Each java class maps to a table. The table lists instances without subclasses

For instance: "java.lang.Object", "java.lang.String" Note: you need to use double quotes to quote identifiers Note: it is assumed that classes sharing the same name have the same fields

The fields become columns.

The following special columns are added:

this         | reference to current object

Fields are available as MAP.get call:

select path -- retrieve field as usual
     , this['path'] -- retrieve field via MAP get
     , this['@className']
     , this['@class']['@classLoader'] -- nested calls work as well
  from java.net.URL

The following virtual properties are available via MAP.get:

@shallow     | shallow heap size of referenced object
@retained    | retained heap size for referenced object
@class       | the same as `getClass()` in Java
@className   | class name (the same as `getClass().getName()` in Java)

The following virtual properties are available for Class instances via MAP.get:

@super       | super class
@classLoader | returns `ClassLoader` for a given class

The following functions can be used to work with column which represents reference:

getId        | internal object identifier for referenced object
getAddress   | memory address for referenced object
getType      | class name of referenced object
toString     | textual representation of referenced object
shallowSize  | shallow heap size of referenced object
retainedSize | retained heap size for referenced object
length       | length of referenced array
getSize      | size of referenced collection, map or count of non-null elements in array
getByKey     | extracts value for given string representation of key for referenced map
getField     | obtains value of field with specified name for referenced object
getStaticField   | obtains value of static field with specified name for referenced class
getStringContent | pretty prints object representation

The following table functions are supported:

getRetainedSet(ref)        | returns the set of retained objects
getOutboundReferences(ref) | returns outbound references (name, this) pairs
getInboundReferences(ref)  | returns inbound references (this)
getValues(ref)             | returns all values of a Java collection
getMapEntries(ref)         | unnests Map as (key, value) tuples

These functions can be called in a following way:

select
 u.this, refs.name, refs.this reference
from 
 java.net.URL u,
 lateral table(getOutboundReferences(u.this)) refs

Another example:

select
 p.this, vals.key, vals."value"
from 
 java.util.Properties p,
 lateral table(getMapEntries(p.this)) vals

The following collection functions are also supported:

asMap(ref)                 | converts Java Map to SQL MAP, so it can be used like asMap(ref)['key']
asMultiSet(ref)            | converts Java Collection to SQL MULTISET type
asArray(ref)               | converts Java references array to SQL ARRAY type
asByteArray(ref)           | converts Java bytes array (bytes[]) to SQL ARRAY type
asShortArray(ref)          | converts Java shorts array (short[]) to SQL ARRAY type
asIntArray(ref)            | converts Java ints array (int[]) to SQL ARRAY type
asLongArray(ref)           | converts Java longs array (long[]) to SQL ARRAY type
asBooleanArray(ref)        | converts Java boolean array (boolean[]) to SQL ARRAY type
asCharArray(ref)           | converts Java chars array (char[]) to SQL ARRAY type
asFloatArray(ref)          | converts Java floats array (float[]) to SQL ARRAY type
asDoubleArray(ref)         | converts Java doubles array (double[]) to SQL ARRAY type

These functions can be called in a following way:

select
 p.this
from 
 java.util.Properties p
where
 asMap(p.this)['java.vm.version'] is not null

Another example:

select 
 fpc.this,
 fp.fp_ref
from 
 java.io.FilePermissionCollection fpc,
 unnest(asMultiSet(fpc.perms)) fp(fp_ref)

Note, that SQL arrays are indexed starting with '1', so following example shows how to get first element of long[] arrays:

select 
 asLongArray(all_long_arrays.this)[1] first_element
from 
 "long[]" all_long_arrays
where 
 length(all_long_arrays.this) > 0

Array index could be extracted into the table by using the 'unnest ... with ordinality' clause:

select
    c.this,
    cs.index,
    cs.val
from
    java.util.GregorianCalendar c,
    unnest(asIntArray(c.stamp)) with ordinality cs(val, index)

Requirements

Java 17 as a build JDK. The code can use all Java 17 features. Eclipse Memory Analyzer 0.14.0 or higher

Building

Eclipse plugin cannot depend on jars from maven repository. It has to be a OSGi bundle, however Calcite is easier to reach via maven. So we use two-phase approach: bundle the dependencies in a single jar, then use this jar in eclipse project.

1. Build the plugin

   ./mvnw install # from the top-level folder
   

   The final repository (aka "update site") with the plugin will be created in MatCalciteRepository/target/MatCalciteRepository-....zip

Running

It is not yet clear how to run the plugin via maven. To launch via Eclipse, just open Eclipse project, double-click plugin.xml, then click Launch an Eclipse application.

Commandline mode

You can process a single SQL via command line as follows

./MemoryAnalyzer -application MatCalcitePlugin.execute <heap-dump.file> <query.file> <result.file>

Roadmap

• enhance heap schema
• context menu support
• heap-related filtering operators (instance of, dominator of, etc)
• interoperation with mat queries (histogram, retained set, etc)
• support external data providers (e.g. allow to join heapdump objects with csv file)
• projection support (do not compute unnecessary columns)
• optimizer rules

License

This library is distributed under terms of Apache 2 License

Author

Vladimir Sitnikov [email protected]