Additional logging questions

[Novice]

I _think_ these supplementary questions will be a lot less of a struggle
than the others in "Aspect questions" thread. That thread is getting on
the long side so I thought I'd ask these in a new thread.

Basically, I'm looking for advice on what should always be logged by
every class. I understand now that every class is going to have its own
logger but what should be logged?

Or to put it another way, are there cases where a class won't log at all?

I'm thinking of things like Enums. If I have an enum that lists the days
of the week, there's not much to go wrong there and I'm not likely to
throw exceptions or even have a try/catch block. So should it just be
left so that it isn't logging at all? Or should there be some standard
bare-minimum sort of logging, like an entering() and existing(), even if
nothing else of interest goes on?

What about holder classes? I'm not sure if I'm using the terminology
correctly but I'm thinking of a class where you simply store related bits
of data, like a Name class whose constructor insists on a first name and
a last name and then supplies getters and setters so that another class
can ask for just the first name or just the last name? (Let's pretend
that everyone has exactly one given name and one surname, no exceptions,
just to keep this simple). This could be an awfully barebones class if it
only had a two line constructor and one line getters and setters. Should
it log anyway?

My feeling is that Lew would say NOT to log unless there was a good
reason to log and then cite several good reasons to log. I'm not sure if
something like an enum or a holder class (if I've used the term
correctly) would EVER justify logging though.

I may have completely misread Lew and, if so, I'm sorry. Maybe this is
another premature leap....

Some of the rest of you may differ dramatically on what should be logged
and when it is okay not to bother. I hope some of you can share those
thoughts with me.

Basically, I'm just about ready to start getting loggers for each and
every class in the project I'm working on now (with plans to do the same
in every project as I create it or return to it). But I don't want to do
too much logging either.....

 

[Arne Vajhøj]

I _think_ these supplementary questions will be a lot less of a struggle
than the others in "Aspect questions" thread. That thread is getting on
the long side so I thought I'd ask these in a new thread.

Basically, I'm looking for advice on what should always be logged by
every class. I understand now that every class is going to have its own
logger but what should be logged?

Or to put it another way, are there cases where a class won't log at all?

I'm thinking of things like Enums. If I have an enum that lists the days
of the week, there's not much to go wrong there and I'm not likely to
throw exceptions or even have a try/catch block. So should it just be
left so that it isn't logging at all? Or should there be some standard
bare-minimum sort of logging, like an entering() and existing(), even if
nothing else of interest goes on?

What about holder classes? I'm not sure if I'm using the terminology
correctly but I'm thinking of a class where you simply store related bits
of data, like a Name class whose constructor insists on a first name and
a last name and then supplies getters and setters so that another class
can ask for just the first name or just the last name? (Let's pretend
that everyone has exactly one given name and one surname, no exceptions,
just to keep this simple). This could be an awfully barebones class if it
only had a two line constructor and one line getters and setters. Should
it log anyway?

My feeling is that Lew would say NOT to log unless there was a good
reason to log and then cite several good reasons to log. I'm not sure if
something like an enum or a holder class (if I've used the term
correctly) would EVER justify logging though.

I may have completely misread Lew and, if so, I'm sorry. Maybe this is
another premature leap....

Some of the rest of you may differ dramatically on what should be logged
and when it is okay not to bother. I hope some of you can share those
thoughts with me.

Basically, I'm just about ready to start getting loggers for each and
every class in the project I'm working on now (with plans to do the same
in every project as I create it or return to it). But I don't want to do
too much logging either.....

You should log the information you expect potentially could be
useful when troubleshooting a problem.

And as a general rule, then if any doubt then log, because it
is usually better to have too much logging than too little
logging.

I do not see any need for logging in an enum or in a pure
data class (holder class).

But please add a toString method in your data class, so
when the class with real login in that uses the data class
can log it and you get something useful in the log about the
data.

Arne

 

[Novice]

You should log the information you expect potentially could be
useful when troubleshooting a problem.

And as a general rule, then if any doubt then log, because it
is usually better to have too much logging than too little
logging.

I do not see any need for logging in an enum or in a pure
data class (holder class).

But please add a toString method in your data class, so
when the class with real login in that uses the data class
can log it and you get something useful in the log about the
data.

Sorry, I'm not following you.

Are you saying that the toString() method needs to be there to turn
things like references into meaningful information? I know that a
reference to something like a JFrame is not going to be very meaningful
and would rather display the name given the JFrame via setName(). Or are
you saying something quite different?

 

[Lew]

You log what reveals to someone reading the log what they want to know.

That depends on the logging level set at runtime, doesn't it?

If a class is unable to log anything, and that omission deprives someone of
necessary information when they're relying on the log, it's a mistake.

If no one ever needs information from that class out of the log, it can get
away with not logging.

Enums are classes. The same considerations apply as for any other class.

Is there? I trust you - there's not.

Good questions. Answer wisely, Grasshopper.

You are.

Logging is generally for state changes.

Don't take advice from your fantasy of me unless it's good advice. Don't take
advice from the real me unless it's good advice, either.

Sure.

Depends on what's in it, doesn't it?

Since I never before said what to log or not log that you've seen, there's
been nothing to misread, has there?

Nothing to read, nothing to misread. It's a simple equation.

This requires that you think like a useful person, not a computer programmer.

When you're troubleshooting a log, you don't have code in front of you. You
have what the log tells you. It had better God-damned tell you what you need,
because you wouldn't be looking if someone weren't breathing down your neck.
No fancy "***********************=============" strings. Logs are dense,
multi-mega- or gigabyte beasts of tightly printed strings.

Ops personnel read logs. Ops personnel think programmers are children. I had
an ops mentor who told me, "We love getting the programmers from [the
development location] here for six months. They go back to coding _changed_!"

Other times they're cursing the programmers who wrote such lame logging
statements.

Some enums.

Like any other class, it depends on what it does. But generally you log state
changes, i.e., behavioral methods (not usually attributes). You log anything
that is weird. You log errors and warnings.

You pick appropriate logging levels. Here's my log4j idiom:

public void loadResource()
{
logger.debug("");

final BufferedReader reader;
try
{
reader = new BufferedReader(new InputStreamReader(getClass()
.getResourceAsStream("/config/configuration.txt")));
}
catch(IOException exc)
{
String msg = "Cannot open configuration. "+ exc.getLocalizedMessage();
logger.error(msg, exc);
throw new IllegalStateException(msg, exc);
}
assert reader != null;

try
{
// read the Reader, etc.
}
catch(IOException exc)
{
String msg = "Cannot read configuration. "+ exc.getLocalizedMessage();
logger.error(msg, exc);
throw new IllegalStateException(msg, exc);
}
finally
{
try
{
reader.close();
}
catch(IOException exc)
{
String msg = "Cannot close configuration. "
+ exc.getLocalizedMessage();
logger.warn(msg, exc);
}
}
}

Note the multiple uses of 'logger' (an instance member) in that method.

Sorry, I'm not following you.

Are you saying that the toString() method needs to be there to turn
things like references into meaningful information? I know that a
reference to something like a JFrame is not going to be very meaningful
and would rather display the name given the JFrame via setName(). Or are
you saying something quite different?

'toString()' should always give a useful way to identify the specific instance.

It should depend on (and usually only on) the same fields used to drive
'hashCode()' and 'equals()' and if supported, 'compareTo()' (which should
always be consistent with each other).

 

[Novice]

You log what reveals to someone reading the log what they want to
know.


That depends on the logging level set at runtime, doesn't it?
Right.

If a class is unable to log anything, and that omission deprives
someone of necessary information when they're relying on the log, it's
a mistake.

Another good rule of thumb.

If no one ever needs information from that class out of the log, it
can get away with not logging.

That seems reasonable to me. I was pretty sure you were not going to
advocate logging for the sake of logging and you didn't let me down ;-)

Enums are classes. The same considerations apply as for any other
class.


Is there? I trust you - there's not.

Well, I'm thinking of enums like mine which are akin to edits. For
instance, I use preferences for some of my programs so I created a
PreferenceTrees enum. It has only two values, SYSTEM and USER. My
PreferenceUtils class uses PreferenceTrees as a type in all of its
methods so that someone invoking those methods can ONLY choose
PreferenceTrees.SYSTEM or PreferenceTrees.USER, there's no possibility
that someone is going to misspell "System", which could certainly happen
if the method was expecting a String parameter. The PreferencesTree enum
is trivial and seems unlikely to ever blow up.

But if an enum did more complicated logic, like looking up a day number
and turning it into a day name in a foreign language, then sure, I can
see that something might go awry and justify logging. The logic that
looks up how to say "Monday" in Latvian or Turkish could throw an
exception that should be logged. But my enums are mostly like the
PreferenceTrees enum that I mentioned.

Good questions. Answer wisely, Grasshopper.


You are.

Some people call these data classes too, if I'm not mistaken. I find that
a bit more descriptive but maybe that's just me.

Logging is generally for state changes.


Don't take advice from your fantasy of me unless it's good advice.
Don't take advice from the real me unless it's good advice, either.

Absolutely. I'm just trying to think along the same lines as you've
proposed in your posts. I hope I haven't made an unwarranted leap again.

Sure.

Depends on what's in it, doesn't it?

See my remarks about enums above. The same would apply to a holder class.
If it's a one line getter or setter, there's probably not much to go
wrong and logging is probably inappropriate. But if it's doing something
that could fail, sure, logging would make sense.

Since I never before said what to log or not log that you've seen,
there's been nothing to misread, has there?

Nothing to read, nothing to misread. It's a simple equation.

Understood. I'm just extrapolating the general principle that you've been
stating, which I would paraphrase (roughly) as "don't do anything
slavishly or unnecessarily; do it because it makes sense to do it".

This requires that you think like a useful person, not a computer
programmer.

When you're troubleshooting a log, you don't have code in front of
you. You have what the log tells you. It had better God-damned tell
you what you need, because you wouldn't be looking if someone weren't
breathing down your neck. No fancy
"***********************=============" strings. Logs are dense,
multi-mega- or gigabyte beasts of tightly printed strings.

Ops personnel read logs. Ops personnel think programmers are children.
I had an ops mentor who told me, "We love getting the programmers from
[the development location] here for six months. They go back to coding
_changed_!"

Other times they're cursing the programmers who wrote such lame
logging statements.

I have had very little contact with operators in the PC era but I did
have some in my mainframe days. Back then, they didn't have a lot to do
with fixing the problems in the sense of repairing the code. Their job
was basically to figure out which program had bombed and then look that
up in their list so they knew which programmer to call.

Would I be right in assuming that it's pretty much the same situation
today in a massively PC-oriented world? Or have they assumed many new
responsibilities?

It would help a lot to know what they hope to find in a log.

If they still operate like they did in my mainframe days, I expect they
pretty much just want to know which program failed so they can look up
the on-call programmer's name. They won't care about most details,
although they might like to be able to tell the programmer "Program Foo
crashed on an IllegalArgumentException in the constructor for class
FooMainPanel" as opposed to just "Program Foo crashed". They won't care
about stacktraces or such things. But the programmer is going to care a
lot about the stacktraces and other information!

But maybe modern operators do a lot more than that. You seem very
familiar with what they do so this would be a great chance to get your
insight on this.

Some enums.

Like any other class, it depends on what it does. But generally you
log state changes, i.e., behavioral methods (not usually attributes).
You log anything that is weird. You log errors and warnings.

You pick appropriate logging levels. Here's my log4j idiom:

public void loadResource()
{
logger.debug("");

final BufferedReader reader;
try
{
reader = new BufferedReader(new InputStreamReader(getClass()
.getResourceAsStream("/config/configuration.txt")));
}
catch(IOException exc)
{
String msg = "Cannot open configuration. "+
exc.getLocalizedMessage(); logger.error(msg, exc);
throw new IllegalStateException(msg, exc);
}
assert reader != null;

try
{
// read the Reader, etc.
}
catch(IOException exc)
{
String msg = "Cannot read configuration. "+
exc.getLocalizedMessage(); logger.error(msg, exc);
throw new IllegalStateException(msg, exc);
}
finally
{
try
{
reader.close();
}
catch(IOException exc)
{
String msg = "Cannot close configuration. "
+ exc.getLocalizedMessage();
logger.warn(msg, exc);
}
}
}

Note the multiple uses of 'logger' (an instance member) in that
method.

Interesting. I'm going to need to imitate that...

'toString()' should always give a useful way to identify the specific
instance.

That's why I was displaying the name of the table so I'm going to take
that as a "yes" ;-)

It should depend on (and usually only on) the same fields used to
drive 'hashCode()' and 'equals()' and if supported, 'compareTo()'
(which should always be consistent with each other).

I don't touch hashCode(), equals() or compareTo() very often at all. Or
toString() either for that matter. But overriding hashCode() and equals()
solved a big problem for me recently. I was writing a holder class that
contained three fields, a table name, a row number and a column number,
as a key for a hash map and then storing the column width in the value
portion of the map. But when I tried to look up a given combination of
table name, row number and column number I was never finding values that
I knew were there. I googled and found out that I needed to revise
hashCode() and equals() and rougly what those changes needed to be and
then my lookups went fine. But I didn't touch compareTo(). Hmm, maybe I
need to revisit that and make sure it doesn't need some tweaking too....

 

[Lew]

I don't touch hashCode(), equals() or compareTo() very often at all. Or
toString() either for that matter. But overriding hashCode() and equals()
solved a big problem for me recently. I was writing a holder class that
contained three fields, a table name, a row number and a column number,
as a key for a hash map and then storing the column width in the value
portion of the map. But when I tried to look up a given combination of
table name, row number and column number I was never finding values that
I knew were there. I googled and found out that I needed to revise
hashCode() and equals() and rougly what those changes needed to be and
then my lookups went fine. But I didn't touch compareTo(). Hmm, maybe I
need to revisit that and make sure it doesn't need some tweaking too....

Read the Javadocs for those methods.

'compareTo()' is not universal, being present only in instances of
'Comparable<T>'.

 

[Daniel Pitts]

I _think_ these supplementary questions will be a lot less of a struggle
than the others in "Aspect questions" thread. That thread is getting on
the long side so I thought I'd ask these in a new thread.

Basically, I'm looking for advice on what should always be logged by
every class. I understand now that every class is going to have its own
logger but what should be logged?

Or to put it another way, are there cases where a class won't log at all?

Classes don't log, methods log. The question then becomes, which methods
should (or shouldn't) log.

What I've determined is that you need to ask yourself a series of questions.

Would logging at this point be useful at all?
Who would it be useful for? Developers, Operations, and/or business?
How much information should I include? Just a message stating an event
happened? Metrics about that event? Details about all objects involved?
What log level should I use?

For example, if you think a Developer needs this information, but no one
else does, then a "debug" level log with all relevant details makes
sense. Operations and/or business probably just want metrics.
Operations just want errors or severe warnings.

Also, depending on where the code runs, you may enable/disable certain
levels of logging. Production probably only warnings/errors should be
enabled after start up, and info should be enabled only during start-up.
In UAT/QA, more information seems useful. Enabling debug makes sense in
DEV environments, *or* when you have a problem you can't reproduce in a
dev environment.

It is also possible (though rare) to have a class have multiple loggers.
I can think of exactly one use-case, but there may be more. That
use-case is specifically for metric logging. If you have some metrics
your code collects to log, then it might make sense to separate those
out into there own "log stream".

Hope this helps,
Daniel.

 

[Lew]

That's why I was displaying the name of the table so I'm going to take
that as a "yes" ;-)

I don't touch hashCode(), equals() or compareTo() very often at all. Or
toString() either for that matter. But overriding hashCode() and equals()
solved a big problem for me recently. I was writing a holder class that
contained three fields, a table name, a row number and a column number,
as a key for a hash map and then storing the column width in the value
portion of the map. But when I tried to look up a given combination of
table name, row number and column number I was never finding values that
I knew were there. I googled and found out that I needed to revise
hashCode() and equals() and rougly what those changes needed to be and
then my lookups went fine. But I didn't touch compareTo(). Hmm, maybe I
need to revisit that and make sure it doesn't need some tweaking too....

The default 'Object#equals()' method gives (almost) the same answer as the ==
operator - two instances are equal iff (if and only if) they are the same
instance. That's called "object equality" or "instance equality".

Often we want value equality. It's not very helpful, for example, for

new Integer(67551) != new Integer(67551)

to be true.

That means equality of different instances, provided their values match in
some defined way.

For 'Integer', value equality depends on only one attribute, the underlying
'int' field.

For a more complicated type, there might be more than one attribute. For
example, you might regard two widgets as equal iff they have the same owner
and color:

public class Widget
{
private final Owner owner; // assume getter methods for brevity
private final Color color;
public Widget(Owner owner, Color color)
{
if (owner == null || color == null)
{
throw new IllegalArgumentException("null");
}
this.owner = owner;
this.color = color'
assert this.owner != null && this.color != null;
}
@Override public boolean equals(Object other)
{
if (other == this) { return true; }
if (! (other instanceof Widget)) { return false; }
Widget widget = (Widget) other;
return this.getOwner().equals(widget.getOwner())
&& this.getColor().equals(widget.getColor()
}
}

But wait! 'hashCode()', which is a shortcut for 'equals()' (well, really for
"not equals()"), doesn't know anything about those fields yet! It will think
objects are not equal when they really are.

So you have to override 'hashCode()', too, using the same fields, so that its
result are consistent with 'equals()':

@Override public int hashCode()
{
return owner.hashCode() * 31 + color.hashCode();
}

Since 'compareTo()' also makes claims about equality, it must agree with the
other two methods when it exists.

Since the fields that determine equality identify the value by definition,
they should determine the 'toString()' result also:

@Override public String toString()
{
assert owner != null && color != null;
return '{'+ owner +", "+ color + '}';
}

As you can see, 'toString()' can include additional information, but it must
show what identifies the value.

 

[Arne Vajhøj]

Sorry, I'm not following you.

Are you saying that the toString() method needs to be there to turn
things like references into meaningful information? I know that a
reference to something like a JFrame is not going to be very meaningful
and would rather display the name given the JFrame via setName(). Or are
you saying something quite different?

I am saying that.

Object toString is not very useful so you want to
override with something that returns relevant information.

Arne

 

[Novice]

Classes don't log, methods log.

You're right of course. I was just shorthanding by talking about classes
logging.

The question then becomes, which
methods should (or shouldn't) log.

What I've determined is that you need to ask yourself a series of
questions.

Would logging at this point be useful at all?
Who would it be useful for? Developers, Operations, and/or business?
How much information should I include? Just a message stating an event
happened? Metrics about that event? Details about all objects
involved? What log level should I use?

For example, if you think a Developer needs this information, but no
one else does, then a "debug" level log with all relevant details
makes sense. Operations and/or business probably just want metrics.
Operations just want errors or severe warnings.

Also, depending on where the code runs, you may enable/disable certain
levels of logging. Production probably only warnings/errors should be
enabled after start up, and info should be enabled only during
start-up. In UAT/QA, more information seems useful. Enabling debug
makes sense in DEV environments, *or* when you have a problem you
can't reproduce in a dev environment.

It is also possible (though rare) to have a class have multiple
loggers.
I can think of exactly one use-case, but there may be more. That
use-case is specifically for metric logging. If you have some metrics
your code collects to log, then it might make sense to separate those
out into there own "log stream".

Hope this helps,

It does! Thanks Daniel!

 

[Arne Vajhøj]

This requires that you think like a useful person, not a computer
programmer.

When you're troubleshooting a log, you don't have code in front of you.
You have what the log tells you. It had better God-damned tell you what
you need, because you wouldn't be looking if someone weren't breathing
down your neck. No fancy "***********************=============" strings.
Logs are dense, multi-mega- or gigabyte beasts of tightly printed strings.

Ops personnel read logs. Ops personnel think programmers are children. I
had an ops mentor who told me, "We love getting the programmers from
[the development location] here for six months. They go back to coding
_changed_!"

Warning and above may be ops relevant, but the lower levels and
that is typical the majority of log messages are for developers.

You pick appropriate logging levels. Here's my log4j idiom:

public void loadResource()
{
logger.debug("");

final BufferedReader reader;
try
{
reader = new BufferedReader(new InputStreamReader(getClass()
.getResourceAsStream("/config/configuration.txt")));
}
catch(IOException exc)
{
String msg = "Cannot open configuration. "+ exc.getLocalizedMessage();
logger.error(msg, exc);
throw new IllegalStateException(msg, exc);
}
assert reader != null;

try
{
// read the Reader, etc.
}
catch(IOException exc)
{
String msg = "Cannot read configuration. "+ exc.getLocalizedMessage();
logger.error(msg, exc);
throw new IllegalStateException(msg, exc);
}
finally
{
try
{
reader.close();
}
catch(IOException exc)
{
String msg = "Cannot close configuration. "
+ exc.getLocalizedMessage();
logger.warn(msg, exc);
}
}
}

Note the multiple uses of 'logger' (an instance member) in that method.

Why do you combine English with local language?

'toString()' should always give a useful way to identify the specific
instance.

It should depend on (and usually only on) the same fields used to drive
'hashCode()' and 'equals()' and if supported, 'compareTo()' (which
should always be consistent with each other).

I would usually expect more fields than the identity fields
to be useful.

Arne

 

[Novice]

The default 'Object#equals()' method gives (almost) the same answer as
the == operator - two instances are equal iff (if and only if) they
are the same instance. That's called "object equality" or "instance
equality".

Often we want value equality. It's not very helpful, for example, for

new Integer(67551) != new Integer(67551)

to be true.

That means equality of different instances, provided their values
match in some defined way.

For 'Integer', value equality depends on only one attribute, the
underlying 'int' field.

For a more complicated type, there might be more than one attribute.
For example, you might regard two widgets as equal iff they have the
same owner and color:

public class Widget
{
private final Owner owner; // assume getter methods for brevity
private final Color color;
public Widget(Owner owner, Color color)
{
if (owner == null || color == null)
{
throw new IllegalArgumentException("null");
}
this.owner = owner;
this.color = color'
assert this.owner != null && this.color != null;
}
@Override public boolean equals(Object other)
{
if (other == this) { return true; }
if (! (other instanceof Widget)) { return false; }
Widget widget = (Widget) other;
return this.getOwner().equals(widget.getOwner())
&& this.getColor().equals(widget.getColor()
}
}

But wait! 'hashCode()', which is a shortcut for 'equals()' (well,
really for "not equals()"), doesn't know anything about those fields
yet! It will think objects are not equal when they really are.

So you have to override 'hashCode()', too, using the same fields, so
that its result are consistent with 'equals()':

@Override public int hashCode()
{
return owner.hashCode() * 31 + color.hashCode();
}

Just curious: why multiply the owner.hashCode() by 31? There's no
significance to the 31 is there? Why not just add the two hashCodes
together without multiplying one of them first? That would be just as
good, right?

Since 'compareTo()' also makes claims about equality, it must agree
with the other two methods when it exists.

Since the fields that determine equality identify the value by
definition, they should determine the 'toString()' result also:

@Override public String toString()
{
assert owner != null && color != null;
return '{'+ owner +", "+ color + '}';
}

As you can see, 'toString()' can include additional information, but
it must show what identifies the value.

That makes perfect sense to me!

 

[Arne Vajhøj]

Arne Vajhøj wrote:
Novice wrote:
Some of the rest of you may differ dramatically on what should be
logged and when it is okay not to bother. I hope some of you can
share those thoughts with me.

Basically, I'm just about ready to start getting loggers for each
and every class in the project I'm working on now (with plans to do
the same in every project as I create it or return to it). But I
don't want to do too much logging either.....

You should log the information you expect potentially could be
useful when troubleshooting a problem.

This requires that you think like a useful person, not a computer
programmer.

When you're troubleshooting a log, you don't have code in front of
you. You have what the log tells you. It had better God-damned tell
you what you need, because you wouldn't be looking if someone weren't
breathing down your neck. No fancy
"***********************=============" strings. Logs are dense,
multi-mega- or gigabyte beasts of tightly printed strings.

Ops personnel read logs. Ops personnel think programmers are children.
I had an ops mentor who told me, "We love getting the programmers from
[the development location] here for six months. They go back to coding
_changed_!"

Other times they're cursing the programmers who wrote such lame
logging statements.

I have had very little contact with operators in the PC era but I did
have some in my mainframe days. Back then, they didn't have a lot to do
with fixing the problems in the sense of repairing the code. Their job
was basically to figure out which program had bombed and then look that
up in their list so they knew which programmer to call.

Would I be right in assuming that it's pretty much the same situation
today in a massively PC-oriented world? Or have they assumed many new
responsibilities?

It would help a lot to know what they hope to find in a log.

If they still operate like they did in my mainframe days, I expect they
pretty much just want to know which program failed so they can look up
the on-call programmer's name. They won't care about most details,
although they might like to be able to tell the programmer "Program Foo
crashed on an IllegalArgumentException in the constructor for class
FooMainPanel" as opposed to just "Program Foo crashed". They won't care
about stacktraces or such things. But the programmer is going to care a
lot about the stacktraces and other information!

But maybe modern operators do a lot more than that. You seem very
familiar with what they do so this would be a great chance to get your
insight on this.

PC's as such does not have operator operators.

But various Unix, Linux and Windows servers have.

How much ops troubleshoot before calling in software
varies between companies.

Some types of problems can be fixed by ops.

But if the problem is really a software bug, then
ops can not do much.

Arne

 

[Arne Vajhøj]

Just curious: why multiply the owner.hashCode() by 31? There's no
significance to the 31 is there? Why not just add the two hashCodes
together without multiplying one of them first? That would be just as
good, right?

Actually not.

Some numbers are better than others (no multiply means 1).

You should go for a prime.

31 is often used.

But there are other good numbers.

Arne

 

[Lew]

Actually not.

Some numbers are better than others (no multiply means 1).

You should go for a prime.

31 is often used.

But there are other good numbers.

Donald Knuth, _The Art of Computer Programming_, suggested 31 for modulo 32
hashes because it produces good distribution on the average for random inputs.

There are better hashes, but for most purposes the one illustrated suffices.

More generally (pseudocode):

@Override int hashCode()
{
int hash = 17; // or 0 or some other decent starter
for (Object field : instanceFields)
{
hash = hash * 31;
hash += field.hashCode();
}
return hash;
}

I skipped over niceties like whether to use 0 for null or empty element hashes.

The Javadocs explain some of this, but the purpose of 'hashCode()' mostly is
to quickly tell if instances are not equal. Equal instances must have the same
hash - that's a hard rule. But the converse cannot apply - many instances will
have the same hash even if they're not equal.

The trick is not to have many all at once during the run of your application.

Then probability is your friend. With 2^^32 possible hashcodes, let's say for
a map of 'String' to 'Foo', you have a lot of room for many strings before you
get too many collisions. That means you need a hash that is about equally
likely to pick any number as any other, so that for any given set of actual
values at a time you probably will have different hashes for different objects.

In practice you get a few collisions, but as long as "a few" is within
acceptable limits you're good to go. Otherwise you have to play advanced
reindeer games with hash algorithms.

Why all this hash fooferol? Because once you have a hash value lookups are so
much faster on 'int' comparisons than between arbitrary 'String's of varying
length. So let's say you're trying to find an element in a 'Set' - you look
for the hashcode first, and if you find zero or one element you're done.
Otherwise you distinguish between the colliding elements with 'equals()', but
hopefully only for two or at worst three elements.

Hashes are also useful in cryptography.

It's an endless topic, full of fun. You need never be bored of a winter's eve.
I'd actually start with Wikipedia on this one.

Aside: I seem to recall an 'equals()' method builder in Apache Commons Lang
somewhere. It's got some sort of indicative name like 'EqualsBuilder'.

 

[Lew]

PC's as such does not have operator operators.

But various Unix, Linux and Windows servers have.

How much ops troubleshoot before calling in software
varies between companies.

Some types of problems can be fixed by ops.

But if the problem is really a software bug, then
ops can not do much.

Sure they can, in partnership with a developer. Someone has to reconfigure
logs to tighter levels (such as DEBUG), restart servers, instantiate test
modes, make logs available to the developer, etc.

As with what you said, this varies between companies.

 

[Arne Vajhøj]

That's one idea I miss in the Java logging idiom: that there are
different types of logs that are kept for different purposes and that
a log message might be relevant for only one or some of these log
types.

That the system can't create a temporary file because there's not
enough available disk space is of vital importance in an operations
log, but completely irrelevant to an audit-trail log and of lesser
importance to an error-tracking / debugging support log.

While it's trivial to implement logging to do this, as the idea hasn't
taken hold in the Java culture it's hardly ever done, which is a pity.

I think the split out of ops relevant data is typical done via the
ops monitoring software (IBM Tivoli, CA Unicenter or whatever).

Arne

 

[Arved Sandstrom]

That's one idea I miss in the Java logging idiom: that there are
different types of logs that are kept for different purposes and that
a log message might be relevant for only one or some of these log
types.

That the system can't create a temporary file because there's not
enough available disk space is of vital importance in an operations
log, but completely irrelevant to an audit-trail log and of lesser
importance to an error-tracking / debugging support log.

While it's trivial to implement logging to do this, as the idea hasn't
taken hold in the Java culture it's hardly ever done, which is a pity.

You're correct on all counts. It _is_ relatively trivial to implement
logging to do this: I worked on logging improvements to several
good-sized apps for a client some years back, and one of the main
features was to build extra classifiers into the formatted log messages.
That way someone could relatively easily Splunk or grep on log entries
that applied to a given subsystem or subsystems.

I suspect one of the main reasons why you don't see more developers do
this is because they get so used to being spoonfed by feature-rich
frameworks that they don't fully realize that there are still big gaps
that they can and should fill. Of all the conversion specifiers provided
by log4j for PatternLayout, for example, it's that innocuous little %m
which is most often not taken advantage of. It's up to the logging
designer to mandate extra formatting information there, such as user IDs
or functional classifiers, to go along with event-specific messages.

I've seen - more often than I care to remember - web app logging that
was unable to support pulling out just the log entries that pertained to
a given authenticated user. Other logs I've seen have no provision to
write out the stack trace of an exception (log4j actually supports this,
but it's amazing how many folks miss that in the API). It makes a person
wonder why people even bother writing to logs.

AHS