BugMakers :)

Frankly speaking, when I was writing previous blogpost I got surprised to discover that DQL update statement preserves the value of r_lock_owner attribute:

API> revert,c,09024be98006b104
...
OK
API> get,c,09024be98006b104,r_lock_owner
...
dmadmin
API> ?,c,update dm_document objects set object_name='xxx' where r_object_id='09024be98006b104'
objects_updated
---------------
              1
(1 row affected)
[DM_QUERY_I_NUM_UPDATE]info:  "1 objects were affected by your UPDATE statement."


API> revert,c,09024be98006b104
...
OK
API> get,c,09024be98006b104,r_lock_owner
...
dmadmin
API> 

Unfortunately, it is not true when you update objects, which behaviour is customized via TBO:

API> retrieve,c,bs_doc_cash
...
09bc2c71806d6ffe
API> checkout,c,l
...
09bc2c71806d6ffe
API> ?,c,update bs_doc_cash objects set object_name='xxx' where r_object_id='09bc2c71806d6ffe'
objects_updated
---------------
              1
(1 row affected)
[DM_QUERY_I_NUM_UPDATE]info:  "1 objects were affected by your UPDATE statement."


API> revert,c,09bc2c71806d6ffe
...
OK
API> get,c,09bc2c71806d6ffe,r_lock_owner
...

API>

but:

API> checkout,c,09bc2c71806d6ffe
...
09bc2c71806d6ffe
API> apply,c,,EXEC,QUERY,S,update bs_doc_cash objects set object_name='xxx' where r_object_id='09bc2c71806d6ffe',BOF_DQL,B,F
...
q0
API> next,c,q0
...
OK
API> dump,c,q0
...
USER ATTRIBUTES

  objects_updated                 : 1

SYSTEM ATTRIBUTES


APPLICATION ATTRIBUTES


INTERNAL ATTRIBUTES


API> revert,c,09bc2c71806d6ffe
...
OK
API> get,c,09bc2c71806d6ffe,r_lock_owner
...
dmadmin
API> 

javassist

Six months ago Marian Caikovski have shared his experience with java decompilers and advertised CFR decompiler: Decompiling jars obfuscated with AspectJ (e.g. D2FS4DCTM-WEB-4.5.0.jar or dfc.jar) – in free clicks from CFR decompiler page (FAQ -> Procyon / Java Decompiler -> Konloch/bytecode-viewer) you may find another cool project indented to provide universal GUI for java decompilers – Konloch/bytecode-viewer:

Well, why is this blogpost named “javassist”? The problem is I do not trust decompilers because some of them produce completely wrong results, for example, I have noticed that a lot of people do like jd-gui decompiler, because it has cool GUI, but try to decompile following code in jd-gui:

public final class Test { 
    public static void main(String[] argv) throws Exception {
        int i = 0;
        while (++i > 0) {
            try {
                if (i < 10) {
                    throw new Exception("xxx");
                }
                System.out.print("xxx");
                break;
            } catch (Exception e) {
                if (!(i < 10)) {
                    throw e;
                }
            }
        }
    }
}

and you will get something like:

public final class Test
{
  public static void main(String[] paramArrayOfString)
    throws Exception
  {
    int i = 0;
    while (true) { i++; if (i <= 0) break;
      try {
        if (i < 10) {
          throw new Exception("xxx");
        }
        System.out.print("xxx");
      }
      catch (Exception localException) {
        if (i >= 10)
          throw localException;
      }
    }
  }
}

which is completely wrong because decompiled code contains infinite loop while original one does not 😦 Moreover, in general, if we want to changefix the behaviour of buggy class we do not need to reconstruct the original source code of the whole class – in the worst case we just need to reconstruct the source code of buggy method, and in the most cases we do not need to reconstruct source code at all, and javassist helps a lot there, let me provide some examples.

When displaying objects in a grid Webtop likes to throw annoying exceptions like:

at com.documentum.fc.client.impl.docbase.DocbaseExceptionMapper.newException(DocbaseExceptionMapper.java:49)
 at com.documentum.fc.client.impl.connection.docbase.MessageEntry.getException(MessageEntry.java:39)
 at com.documentum.fc.client.impl.connection.docbase.DocbaseMessageManager.getException(DocbaseMessageManager.java:137)
 at com.documentum.fc.client.impl.connection.docbase.netwise.NetwiseDocbaseRpcClient.checkForMessages(NetwiseDocbaseRpcClient.java:310)
 at com.documentum.fc.client.impl.connection.docbase.netwise.NetwiseDocbaseRpcClient.applyForObject(NetwiseDocbaseRpcClient.java:653)
 at com.documentum.fc.client.impl.connection.docbase.DocbaseConnection$8.evaluate(DocbaseConnection.java:1370)
 at com.documentum.fc.client.impl.connection.docbase.DocbaseConnection.evaluateRpc(DocbaseConnection.java:1129)
 at com.documentum.fc.client.impl.connection.docbase.DocbaseConnection.applyForObject(DocbaseConnection.java:1362)
 at com.documentum.fc.client.impl.docbase.DocbaseApi.parameterizedFetch(DocbaseApi.java:107)
 at com.documentum.fc.client.impl.objectmanager.PersistentDataManager.fetchFromServer(PersistentDataManager.java:201)
 at com.documentum.fc.client.impl.objectmanager.PersistentDataManager.getData(PersistentDataManager.java:92)
 at com.documentum.fc.client.impl.objectmanager.PersistentObjectManager.getObjectFromServer(PersistentObjectManager.java:355)
 at com.documentum.fc.client.impl.objectmanager.PersistentObjectManager.getObject(PersistentObjectManager.java:311)
 at com.documentum.fc.client.impl.session.Session.getObject(Session.java:946)
 at com.documentum.fc.client.impl.session.SessionHandle.getObject(SessionHandle.java:652)
 at com.documentum.webcomponent.library.actions.ExportAction.queryExecute(Unknown Source)
 at com.documentum.web.formext.action.ActionService.queryExecute(Unknown Source)
 at com.documentum.web.formext.control.action.ActionMultiselect.getMultiselectActionExecuteTable(ActionMultiselect.java:615)
 at com.documentum.web.formext.control.action.ActionMultiselectTag.renderEnd(ActionMultiselectTag.java:185)
 at com.documentum.web.form.ControlTag.doEndTag(Unknown Source)

the root cause of such behaviour is following: webtop selects a list of objects to display and after that it calculates preconditions, the problem is when webtop calculates preconditions the list of objects is already not actual – somebody might delete some objects or change permissions and, hence, some preconditions throw an exception. How to fix this behaviour? I do think it is obvious (but not for talented team) that queryExecute method of ActionService class must return false instead of throwing exceptions, but if we try to decompile ActionService class we will find that it contains about 500 lines of code and, moreover, decompilers do not produce a “beautiful” code – it will take a couple of hours to reconstruct original source code of ActionService class, meanwhile the javassist solution looks extremely robust and simple:

public class ActionServiceFix {

    public static void main(String[] args) throws Exception {
        ClassPool pool = ClassPool.getDefault();
        for (Class cls : new Class[] {Trace.class }) {
            pool.importPackage(cls.getPackage().getName());
        }

        CtClass cc = pool.get(ActionService.class.getName());
        for (Class cls : new Class[] {DfException.class }) {
            pool.importPackage(cls.getPackage().getName());
        }
        CtMethod queryExecute = cc.getDeclaredMethod("queryExecute",
                pool.get(new String[] {ActionService.ActionDef.class.getName(),
                    ArgumentList.class.getName(), Context.class.getName(),
                    Component.class.getName(), boolean.class.getName() }));
        queryExecute.addCatch("Trace.error(null,t);\n;return false;",
                pool.get(Throwable.class.getName()), "t");
        File f = File.createTempFile(ActionService.class.getSimpleName(), ".class");
        FileOutputStream os = new FileOutputStream(f);
        os.write(cc.toBytecode());
        os.close();
        System.out.println(f.getAbsolutePath());
    }

}

Another concurrency issue I already complained about: implementation of aspects in DFC causes DM_OBJ_MGR_E_UNABLE_TO_FETCH_CONSISTENT_OBJECT_SNAPSHOT errors:

public class DfcFix {

    public static void main(String[] args) throws Exception {
        ClassPool pool = ClassPool.getDefault();
        for (Class cls : new Class[] {}) {
            pool.importPackage(cls.getPackage().getName());
        }

        String POM = PersistentObjectManager.class.getName().replace(".", "/")
                + ".class";
        URL resourceURL = DfcFix.class.getResource("/" + POM);
        JarURLConnection connection = (JarURLConnection) resourceURL
                .openConnection();
        URL jarURL = connection.getJarFileURL();
        String fileName = jarURL.getFile();
        ZipFile zipFile = new ZipFile(fileName);
        String out = fileName + ".patched";
        ZipOutputStream zos = new ZipOutputStream(new FileOutputStream(out));
        for (Enumeration e = zipFile.entries(); e.hasMoreElements();) {
            ZipEntry entryIn = (ZipEntry) e.nextElement();
            if (entryIn.getName().equals(POM)) {
                zos.putNextEntry(new ZipEntry(POM));
                zos.write(getPOMFixRename(pool));
            } else {
                zos.putNextEntry(new ZipEntry(entryIn.getName()));
                InputStream is = zipFile.getInputStream(entryIn);
                byte[] buf = new byte[1024];
                int len;
                while ((len = is.read(buf)) > 0) {
                    zos.write(buf, 0, len);
                }
            }
            zos.closeEntry();
        }
        zos.close();
        System.out.println(out);
    }

    private static byte[] getPOMFixRename(ClassPool pool) throws Exception {
        CtClass cc = pool.get(PersistentObjectManager.class.getName());
        for (Class cls : new Class[] {DfException.class, DfExceptions.class, }) {
            pool.importPackage(cls.getPackage().getName());
        }

        CtMethod original = cc.getDeclaredMethod("getObject");
        CtMethod newMethod = CtNewMethod.make(AccessFlag.PUBLIC,
                original.getReturnType(), original.getName(),
                original.getParameterTypes(), original.getExceptionTypes(),
                null, original.getDeclaringClass());
        original.setName("doGetObject");
        original.setModifiers(AccessFlag.setPrivate(original.getModifiers()));
        StringBuilder body = new StringBuilder();
        body.append("{\n");
        body.append("DfException ex = null;\n");
        body.append("for (int i = 0; i < 10; i++) {\n");
        body.append("    try {\n");
        body.append("        return doGetObject($$);\n");
        body.append("    } catch (DfException e) {\n");
        body.append("        ex = e;\n");
        body.append("        if (DfExceptions.isFetchSoft(e)) {\n");
        body.append("            DfLogger.debug(this, \"Got soft exception \"\n");
        body.append("                    + \"on {0} iteration\", new Object[] {i + 1 }, e);\n");
        body.append("            continue;\n");
        body.append("        }\n");
        body.append("        throw ex;\n");
        body.append("    }\n");
        body.append("}\n");
        body.append("throw ex;\n");
        body.append("\n}");
        newMethod.setBody(body.toString());
        cc.addMethod(newMethod);
        return cc.toBytecode();

    }

}

Beware of dbi services

Do you remember a guy, who accidentally discovered SQL injection in Content Server? I can’t understand why some people do such things, so I take it for granted that we can’t prevent such misbehaviour, however I wonder why these people come up with heart-piercing stories. Below are two another stories:

Documentum – Not able to install IndexAgent with xPlore 1.6 – everything is good except following command listing:

[xplore@full_text_server_01 ~]$ echo 'export DEVRANDOM=/dev/urandom' >> ~/.bash_profile
[root@full_text_server_01 ~]# yum -y install rng-tools.x86_64
Loaded plugins: product-id, search-disabled-repos, security, subscription-manager
Setting up Install Process
Resolving Dependencies
--> Running transaction check
...
Transaction Test Succeeded
Running Transaction
  Installing : rng-tools-5-2.el6_7.x86_64                                                                                                                                                                                     1/1
  Verifying  : rng-tools-5-2.el6_7.x86_64                                                                                                                                                                                     1/1
 
Installed:
  rng-tools.x86_64 0:5-2.el6_7
 
Complete!
[root@full_text_server_01 ~]# rpm -qf /etc/sysconfig/rngd
rng-tools-5-2.el6_7.x86_64
[root@full_text_server_01 ~]#
[root@full_text_server_01 ~]# sed -i 's,EXTRAOPTIONS=.*,EXTRAOPTIONS=\"-r /dev/urandom -o /dev/random -t 0.1\",' /etc/sysconfig/rngd
[root@full_text_server_01 ~]# cat /etc/sysconfig/rngd
# Add extra options here
EXTRAOPTIONS="-r /dev/urandom -o /dev/random -t 0.1"
[root@full_text_server_01 ~]#
[root@full_text_server_01 ~]# chkconfig --level 345 rngd on
[root@full_text_server_01 ~]# chkconfig --list | grep rngd
rngd            0:off   1:off   2:off   3:on    4:on    5:on    6:off
[root@full_text_server_01 ~]#
[root@full_text_server_01 ~]# service rngd start
Starting rngd:                                             [  OK  ]
[root@full_text_server_01 ~]#

which actually looks exactly the same as my recommendations for increasing entropy on Linux/VMWare, and the real gem is how blogpost author tried to protect himself – there are even four explanations why it looks extremely similar:

  • I would say the source is myself
  • At that time, I opened a SR# with the EMC Support
  • These commands haven’t been provided by EMC, they are part of our IQs since 2014/2015
  • Moreover how is that a proof? I mean all I did is a sed command to update the file /etc/sysconfig/rngd and the setup of the rngd service using chkconfig… There is no magic here, there is nothing secret…

Well, I would buy the last explanation if there were no following inconsistencies:

  • What was the reason to execute rpm -qf /etc/sysconfig/rngd if you already installed rng-tools? In my recommendations I used this command to show where /etc/sysconfig/rngd file came from
  • DEVRANDOM environment variable affects Content Server only, in java environment it does not make sense
  • The second blogpost, see below…

Documentum – Increase the number of concurrent_sessions – initially the solution was posted 4 years ago on ECN blog, moreover it is also published in EMC KB (note the publication date – it is not consistent with “A few months ago at one of our customer …” statement):

and in another EMC KB (wow! there is a mention of 1100):

Actually, as it was mentioned in my ECN blogpost – the DM_FD_SETSIZE “option” is “officially” available since 6.7SP1P19 and 6.7SP2P04 (and as well in 7.0, 7.1, 7.2 and 7.3, not officially this option is available since 6.7SP1P15), so, I wonder how it was possible that DBI guys were able to do following:

An EMC internal task (CS-40186) was opened to discuss this point and to discuss the possibility to increase this maximum number. Since the current default limit is set only in regards to the default OS value of 1024, if this value is increased to 4096 for example (which was our case since the beginning), then there is no real reason to be stuck at 1020 on Documentum side. The Engineering Team implemented a change in the binaries that allows changing the limit

Moreover, there is another inconsistency: until CS-40517 EMC was suggesting to launch multiple Content Server instances on the same host in order to overcome the limit on 1020 concurrent sessions per Content Server instance, so in case of blogpost author he was need to launch two Content Servers on each host and get an overall limit of 4080 concurrent sessions, but in my case I was need to launch about 10 Content Servers, and, because I was considering such configuration as unmanageable, I performed some research and filed a CR on November 2012.

Database connections

Let’s discuss another statement from performance guide (forget about the fact that the writer experiences difficulties with math (I do believe that 2x2x250 is 1000 but not 2000)):

Tune PROCESSES: As a rule of thumb, PROCESSES should be set to the double of the sum of maximum concurrent sessions of all content server instances. For example, if there are two instances each with concurrent_sessions set to 250, PROCESSES should be set to 2000 (2x2x250). It needs to specify the scope in the alter command when change the value.

Well, the question is: why does docbase session require two database connections? Actually, it is worth writing a continuation for “r_object_id. Type tag” blogpost and describe how Content Server (actually DFC) generates identifiers, but the common idea is following: when DFC creates new object (i.e. when we call IDfSession#newObject method) it sends NEXT_ID or NEXT_ID_LIST RPC command to Content Server:

API> apply,c,,NEXT_ID_LIST,TAG,I,8,HOW_MANY,I,10
...
q0
API> ?,c,q0
next_id         
----------------
08024be98005fe92
08024be98005fe93
08024be98005fe94
08024be98005fe95
08024be98005fe96
08024be98005fe97
08024be98005fe98
08024be98005fe99
08024be98005fe9a
08024be98005fe9b
(1 row affected)

and Content Server modifies data stored in dmi_sequence_s table. Now imagine that we are creating tons of objects in parallel and, moreover, we are taking advantage of using transactions, how do NEXT_ID/NEXT_ID_LIST commands work in this case? If CS modifies data in dmi_sequence_s table in transaction, corresponding rows become locked and another transaction unable to modify the same data, so, to prevent such concurrency issue when performing NEXT_ID/NEXT_ID_LIST commands Content Server creates new (temporary) database connection and never releases it:

API> ?,c,execute EXEC_SELECT_SQL with QUERY='SELECT COUNT(*) cnt 
     FROM v$session WHERE USERNAME=SYS_CONTEXT(''USERENV'',''SESSION_USER'')'
cnt                   
----------------------
                    14
(1 row affected)

API> apply,c,,NEXT_ID_LIST,TAG,I,8,HOW_MANY,I,1000
...
q0
API> ?,c,execute EXEC_SELECT_SQL with QUERY='SELECT COUNT(*) cnt 
     FROM v$session WHERE USERNAME=SYS_CONTEXT(''USERENV'',''SESSION_USER'')'
cnt                   
----------------------
                    15
(1 row affected)

you may observe the reference to this temporary database connection in output of SHOW_SESSIONS command (note the value of tempdb_session_ids, also note that, in general, tempdb_session_ids in SHOW_SESSIONS output displays wrong data):

USER ATTRIBUTES

  root_start                      : 5/10/2017 09:17:05
  root_pid                        : 2502
  shared_mem_id                   : 7667716
  semaphore_id                    : 327685
  session                      [0]: 01024be9800060d7
  db_session_id                [0]: 341
  typelockdb_session_id        [0]: -1
  tempdb_session_ids           [0]: 328
  pid                          [0]: 21067
  user_name                    [0]: dmadmin
  user_authentication          [0]: Trusted Client
  client_host                  [0]: docu72dev01
  client_lib_ver               [0]: 7.2.0030.0072
  client_locale                [0]: (Linux :(8201), Version: Linux), CharSet: UTF-8, Language: English_US, UTC Offset: 14400, Date Form
  start                        [0]: 5/10/2017 13:33:16
  last_used                    [0]: 5/10/2017 13:36:52
  session_status               [0]: Active
  shutdown_flag                [0]: none
  last_rpc                     [0]: EXEC, EXEC, NEXT_ID_LIST, EXEC, EXEC, TIME, FETCH_TYPE, GET_SERVER_CONFI, GET_ERRORS, AUTHENTICATE_USE, ENTRY_POINTS,
  current_rpc                  [0]: SHOW_SESSIONS
  last_completed_rpc           [0]: 5/10/2017 13:34:39
  transaction_status           [0]: Off
  dormancy_status              [0]: Active

So, in long term perspective (actually, both CS and DFC tries to pool “available” identifiers) every docbase session tends to have at least two associated database connections – that is a starting point for database capacity planning, but the question is whether it is an upper bound or not, i.e if we follow “best practices” and set the value of PROCESSES parameter to the double of the sum of maximum concurrent sessions of all content server instances will it prevent us from getting “ORA-00020: maximum number of processes (%s) exceeded” errors or not? And, unfortunately, the answer is: no, it won’t:

API> execquery,c,F,select * from dm_user
...
OK
API> execquery,c,F,select * from dm_user
...
OK
API> execquery,c,F,select * from dm_user
...
OK
API> execquery,c,F,select * from dm_user
...
OK
API> execquery,c,F,select * from dm_user
...
OK
API> apply,c,,SHOW_SESSIONS
...
q7
API> next,c,q7
...
OK
API> dump,c,q7
...
USER ATTRIBUTES

  root_start                      : 5/10/2017 09:17:05
  root_pid                        : 2502
  shared_mem_id                   : 7667716
  semaphore_id                    : 327685
  session                      [0]: 01024be9800060d7
  db_session_id                [0]: 341
  typelockdb_session_id        [0]: -1
  tempdb_session_ids           [0]: 328:74:352:281:270:141
  pid                          [0]: 21787
  user_name                    [0]: dmadmin
  user_authentication          [0]: Trusted Client
  client_host                  [0]: docu72dev01
  client_lib_ver               [0]: 7.2.0030.0072
  client_locale                [0]: (Linux :(8201), Version: Linux), CharSet: UTF-8, Language: English_US, UTC Offset: 14400, Date Form
  start                        [0]: 5/10/2017 13:45:30
  last_used                    [0]: 5/10/2017 13:45:37
  session_status               [0]: Active
  shutdown_flag                [0]: none
  last_rpc                     [0]: EXEC, EXEC, EXEC, EXEC, EXEC, GET_ERRORS, AUTHENTICATE_USE, ENTRY_POINTS, 
  current_rpc                  [0]: SHOW_SESSIONS
  last_completed_rpc           [0]: 5/10/2017 13:45:34
  transaction_status           [0]: Off
  dormancy_status              [0]: Active

CMIS

About 6 moths ago I was complaining about CMIS that it considers all requests which contain the same credentials (i.e. login and password are the same for all requests) as requests from the same client, and for all requests containing the same credentials CMIS uses the only one repository session, unfortunately we didn’t implement solution proposed in that blogpost – it is hard to maintain different passwords across clients, meanwhile we have started receiving concurrency-related errors, so, it was required to undertake something and I have found a solution – it is enough to replace just two classes in CMIS:

dfc.query.should_include_object_name

Have never thought that my colleagues may teach me something…

Yesterday I asked my colleague, who is trying to improve his skills in performance optimisation, whether he had any idea how to improve this SQL statement:

SELECT ALL dm_folder.r_object_id
  FROM dm_folder_sp dm_folder
 WHERE     (    EXISTS
                   (SELECT r_object_id
                      FROM dm_folder_r
                     WHERE     dm_folder.r_object_id = r_object_id
                           AND r_folder_path = :"SYS_B_00")
            AND (dm_folder.object_name = :"SYS_B_01"))
       AND (    dm_folder.i_has_folder = :"SYS_B_02"
            AND dm_folder.i_is_deleted = :"SYS_B_03")

and, surprisingly, the answer was: “Yes, I have seen something similar on support site – EMC suggest to set dfc.query.should_include_object_name and dfc.query.should_include_object_name properties, something like:

dfc.query.object_name_for_docbase[0]=<docbase_name>
dfc.query.should_include_object_name[0]=false


Well, as was expected both dfc.query.should_include_object_name and dfc.query.should_include_object_name properties are not documented, so let discuss the problem more thoroughly.

Imagine that we are maintaining following folder structure in our docbase:

\_CLIENT_1
  \_CLAIMS
  \_INVOCES
\_CLINET_2
  \_CLAIMS
  \_INVOCES
...
\_CLIENT_X
  \_CLAIMS
  \_INVOCES

i.e. for every client we create the same folder structure and when we want to store invoice for particular client we do something like:

create,c,dm_document
set,c,l,object_name
xxx
link,c,/CLIENTS/CLINET_1/INVOICES
save,c,l

the problem is that upon link call DFC calls IDfSession#getFolderByPath method to retrieve folder object with particular path, and inside IDfSession#getFolderByPath method DFC does following: it cuts off object name part from the path (i.e. everything after last ‘/’) and builds following DQL query:

SELECT r_object_id FROM dm_folder 
WHERE object_name='INVOICES' 
 AND ANY r_folder_path='/CLIENTS/CLINET_1/INVOICES'

such implementation is bit weird for two reasons:

  • when I do the same I just write something like “retrieve,c,dm_folder where any r_folder_path=”…” and do not bother myself about object name
  • Content Server has a build-in FolderIdFindByPath RPC command:
    API> apply,c,,FolderIdFindByPath,_FOLDER_PATH_,S,/dmadmin
    ...
    q0
    API> next,c,q0
    ...
    OK
    API> get,c,q0,result
    ...
    0c01d92080000105
    API> close,c,q0
    ...
    OK

    which generates following effective SQL statement:

    select r_object_id from dm_folder_r where r_folder_path = :p0

so, I have no idea why DFC performs extra logic here, moreover, in case of current DFC implementation we are getting overcomplicated SQL query and, sometimes database engine fails to build a good execution plan for this query (this is caused by dumb recommendation to set CURSOR_SHARING database parameter to FORCE and depending on docbase structure execution of such query may take minutes). Below are two possible execution plans for this query:

good (dm_folder_r is a leading table – querying dm_folder_r table by r_folder_path will always return not more than one row):

and bad (dm_folder_r is not a leading table – imagine that we have 1 million clients and hence 1 million INVOICE folders, so querying dm_sysobjec_s table by object_name first will return 1 million records):

in case of “retrieve,c,dm_folder where any r_folder_path=”…” execution plan is always good:

In 2011 (if my memory serves me right), I solved such performance problem by marking index on dm_folder_r(r_folder_path) as unique – in this case database engine always builds the correct execution plan because it knows that querying dm_folder_r table will always return not more than one row, however in recent versions DFC it is possible to disable it’s dumb behaviour by setting dfc.query.should_include_object_name and dfc.query.should_include_object_name properties – can’t understand why this wasn’t enabled by default.

anti-performance series

It is been already a year since I had started nurturing an idea how to write a blogpost about performance best practices, unfortunately, such idea was initially doomed to failure – there are a lot of materials that need to be verified before posting, and no doubt it should take a lot of time, so I “invented” another format: I will try to prove or ruin statements from performance guides provided by talented team.

Actually, some performance-related statements were already ruined in previous posts:

Minimizing and consolidating activities
System throughput varies between 3-100 activities per second, depending on system configuration and hardware. Workflows with more activities take longer to complete. The largest performance impact for processing activities results from opening up a new Content Server session. As a result, the biggest performance improvement comes from minimizing the number of discrete activities in a workflow. Minimize the number of workflow activities by, 1) eliminating unnecessary activities altogether or 2) consolidating the steps performed by multiple activities, into a single condensed activity.
To improve the completion rate of individual activities, do the following:

  • Use the bpm_noop template wherever possible. This particular noop does not create an additional template and does not send an HTTP post to the JMS
  • Within the automatic activity, do the work on behalf of a superuser instead of a regular user
  • Turn off auditing whenever unnecessary

Iteratively modify the number of system workflow threads to assess the impact on user response time, activity throughput, and system resource consumption. More workflow threads result in greater automatic activity throughput up to the point where system resource consumption degrades performance. Scale up slowly to understand when resource limitations begin to show (Content Server CPU and database CPU utilization). The following provides some guidelines:

  • A single CPU Content Server host cannot process 10,000 activities per hour, regardless of how it is configured
  • Be cautious if CPU or memory utilization exceeds 80% for any tier in the system
  • Do not configure more than three threads per CPU core

If throughput requirements exceed the capacity that a single Content Server can provide, add more Content Servers. Each Content Server instance (and associated workflow agent) nominally supports 15 concurrent workflow threads. Deploy one Content Server instance for every multiple of 15 concurrent workflow threads required by your solution. Avoid more than 25 workflow threads for any Content Server.

In general the statements above are misleading:

  • I doubt that “The largest performance impact for processing activities results from opening up a new Content Server session”: at first, JMS do not open new sessions – all sessions are already in session pool, bad thing here is DFC performs authentication when it acquires session from session pool – CS generates new ticket for every auto-activity and these tickets never match passwords associated with pooled sessions and, if my memory severs me right, such reauthentication takes 2 RPCs, at second, dealing with workitem typically takes 4 RPCs: begin transaction, acquire, complete, commit (+ content server does some extra job: creating next activity, updating workflow object, etc) + we need to do some useful work (i.e. perform business logic)
  • workflow delays, caused by processing of auto-activities, does not affect business users: business users are not robots, they do not complete tasks as quick as thought – a couple of extra minutes won’t make sense. On the other hand “consolidating” auto-activities has a negative impact on a project complexity: you need to either consolidate both code and docbase methods or create an extra layer, purposed to implement such consolidation (actually, we use the second option, but that wasn’t influenced by performance considerations), so, it is much better to keep code simple in spite of EMC’s idea about consolidations sounds reasonable
  • I have no idea what were prerequisites to suggest invoking auto-activities under superuser account (I would accept the following scenario: all auto-activities are invoked under installation owner account and CS/JMS takes advantage of trusted authentication, but workflow agent does not support such option), but my preferred option is to assign “previous activity performer” as performer of auto-activity and take advantage of dynamic groups – such approach allows to keep track of last performer of manual activities – business users are able to see who have sent them a task
  • “10,000 auto-activities per hour for single CPU host” is extremely pessimistic estimation – 30,000-50,000 is more close to reality on modern hardware
  • there is no scientific explanation why we need to limit the number of workflow agents by 25 (extra licence fees?) – I do believe that “2 * number of cores” is a good starting point for any hardware configuration