Despite the best efforts of software engineers to produce high-quality software, inevitably some bugs escape even the most rigorous testing process and are first encountered by end users. When this happens, such failures must be understood quickly, the underlying bugs fixed, and deployments patched to avoid another user (or the same one) running into the same problem again. As far back as 1951, the dawn of modern computing, Stanley Gill6 wrote that "some attention has, therefore, been given to the problem of dealing with mistakes after the programme has been tried and found to fail." Gill went on to describe the first use of "the post-mortem technique" in software, whereby the running program was modified to record important system state as it ran so that the programmer could later understand what happened and why the software failed.
Since then, postmortem debugging technology has been developed and used in many different systems, including all major consumer and enterprise operating systems, as well as the native execution environments on those systems. These environments make up much of today's core infrastructure, from the operating systems that underlie every application to core services such as DNS (Domain Name System), and thus form the building blocks of nearly all larger systems. To achieve the high levels of reliability expected from such software, these systems are designed to restore service quickly after each failure while preserving enough information that the failure itself can later be completely understood.
Conspicuously absent from many of these environments, however, are facilities for even basic postmortem debugging, which makes understanding production failures extremely difficult. Dynamic languages must bridge this gap and provide rich tools for understanding failures in deployed systems in order to match the reliability demanded from their growing role in the bedrock of software systems.
To understand the real potential for sophisticated postmortem analysis tools, we will first review the state of debugging today and the role of postmortem analysis tools in other environments. We will then examine the unique challenges around building such tools for dynamic environments and the state of such tools today.
To understand the unique value of postmortem debugging, it's worth examining the alternative. Both native and dynamic environments today provide facilities for in situ debugging, or debugging faulty programs while they're still running. This typically involves attaching a separate debugger program to the faulty program and then directing execution of the faulty program interactively, instruction by instruction or using breakpoints. The user thus stops the program at various points to inspect program state in order to figure out where it goes wrong. Often this process is repeated to test successive theories about the problem.
This technique can be very effective for bugs that are readily reproducible, but it has several drawbacks. First, the act of stopping a program often changes its behavior. Bugs resulting from unexpected interactions between parallel operations (such as race conditions) can be especially challenging to analyze this way because the timing of various events is significantly affected by the debugger itself. More importantly, in situ debugging is often untenable on production systems: many debuggers rely on an unoptimized debug build that's too slow to run in production; engineers often do not have access to the systems where the program is running (as in the case of most mobile and desktop applications and many enterprise systems); and the requisite debugging tools are often not available on those systems anyway.
Even for cases where engineers can access the buggy software with the tools they need, pausing the program in the debugger usually represents an unacceptable disruption of production service and an unacceptable risk that a fat-fingered debugger command might cause the program to crash. Administrators often cannot take the risk of downtime now in order to understand a failure that caused a previous outage. More importantly, they should not have to. Even in 1951 Gill cited the "extravagant waste of machine time involved" in concluding that "single-[step] operation is a useful facility for the maintenance engineer, but the programmer can only regard it as a last resort."
The most crippling problem with in situ debugging is that it can only be used to understand reproducible problems. Many production issues are either very rare or involve complex interactions of many systems, which are often very difficult to replicate in a development environment. The rarity of such issues does not make them unimportant: quite the contrary, an operating system crash that happens only once a week can be extremely costly in terms of downtime, but any bug that can be made to occur only once a week is very hard to debug live. Similarly, a fatal error that occurs once a week in an application used by thousands of people may result in many users hitting the bug each day, but engineers cannot attach a debugger on every user's system.
printf debugging is a common technique for dealing with the reproducibility issue. In this approach, engineers modify the software to log bits of relevant program state at key points in the code. This causes data to be collected without human intervention so that it can be examined after a problem occurs to understand what happened. By automating the data collection, this technique usually results in significantly less impact to production service because when the program crashes, the system can immediately restart it without waiting for an engineer to log in and debug the problem interactively.
Extracting enough information about fatal failures from a log file is often very difficult, however, and frequently it's necessary to run through several iterations of inserting additional logging, deploying the modified program, and examining the output. This, too, is untenable for production systems since ad-hoc code changes are often impractical (in the case of desktop and mobile applications) or prohibited by change control policies (and common sense).
The solution is to build a facility that captures all program state when the program crashes. In 1980 Douglas R. McGregor and Jon R. Malone9 of the University of Strathclyde in Glasgow observed that with this approach "there is virtually no runtime overhead in either space or speed" and "no extra trace routines are necessary," but the facility "remains effective when a program has passed into production use." Most importantly, after the system saves all the program state, it can restart the program immediately to restore service quickly. With such systems in place, even rare bugs can often be root-caused and fixed based on the first occurrence, whether in development, test, or production. This enables software vendors to fix bugs before too many users encounter them.
To summarize, in order to root-cause failures that occur anywhere from development to production, a postmortem debugging facility must satisfy several constraints:
* Application software must not require modifications that cannot be used in production in order to support postmortem debugging, such as unoptimized code or additional debug data that would significantly impact performance (or affect correctness at all).
* The facility must be always on: it must not require an administrator to attach a debugger or otherwise enable postmortem support before the problem occurs.
* The facility must be fully automatic: it should detect the crash, save program state, and then immediately allow the system to restart the failed component to restore service as quickly as possible.
* The dump (saved state) must be comprehensive: a stack trace, while probably the single most valuable piece of information, very often does not provide sufficient information to root-cause a problem from a single occurrence. Usually engineers want both global state and each thread's state (including stack trace and each stack frame's arguments and variables). Of course, there's a wide range of possible results in this dimension; the "constraint" (such as it is) is that the facility must provide enough information to be useful for nontrivial problems. The more information that can be included in the dump, the more likely engineers will be able to identify the root cause based on just one occurrence.
* The dump must be transferable to other systems for analysis. This allows engineers to analyze the data using whatever tools they need in a familiar environment and obviates the need for engineers to access production systems in many cases.
To understand the potential value of postmortem debugging in dynamic languages, it's also helpful to examine those areas where postmortem analysis techniques are well developed and widely used. The best examples are operating systems and their native execution environments. Historically this software has comprised core infrastructure; failures at this level of the stack are often very costly either because the systems themselves are necessary for business-critical functions (as in the case of operating systems on which business-critical software is running) or because they are relied upon by business-critical systems upstack (as in the case of infrastructure services such as DNS).
Most modern operating systems can be configured so that when they crash, they immediately save a "crash dump" of all of their state and then reboot. Likewise, these systems can be configured so that when a user application crashes, the operating system saves a "core dump" of all program state to a file and then restarts the application. In most cases, these mechanisms allow the operating system or user application to return to service quickly while preserving enough information to root-cause the failure later.
As an example, let's look at core dumps under Illumos, an open source Solaris-based system. Take the following broken program:
2 main(int argc, char *argv)
5 return (0);
11 int ii;
12 int *ptrs;
14 for (ii = -1; ii < 100; ii++)
15 *(ptrs[ii]) = 0;
17 return (0);
This simple program has a fatal flaw: while trying to clear each item in the
ptrs array at lines 14-15, it clears an extra element before the array (where
ii = -1). When running this program, you see
and the system generates a file called
core. The Illumos MDB (modular debugger) can help in examining this file:
MDB's syntax may seem arcane to new users, but this example is rather basic. First the
::status command produces a summary of what happened: the process was terminated as a result of a segmentation fault attempting to access memory address 0x10. Next the
::walk thread | ::findstack -v command is used to examine thread stacks (in this case, just one), and it shows that the program died in function
func at offset 0x20 in the program text. Then the file dumps out this instruction to see that the process died on the store of 0 into the address contained in register %eax.
While this example is admittedly contrived, it illustrates the basic method of postmortem debugging. Note that unlike in situ debugging, this method scales well with the complexity of the program being debugged. If instead of one thread in one process there were thousands of threads across dozens of components (as in the case of an operating system), a comprehensive dump would include information about all of them. The next challenge would be making sense of so much information, but root-causing the bug is at least tractable because all the information is available.
In such situations, the next step is to build custom tools for extracting, analyzing, and summarizing specific component state. A comprehensive postmortem facility enables engineers to build such tools. For example, gdb supports user-defined macros. These macros can be distributed with the source code so that all developers can use them both in situ (by attaching gdb to a running process) and postmortem (by opening a core file with gdb). The Python interpreter, for example, provides such macros, allowing both interpreter and native module developers to pick apart the C representations of Python-level objects.
MDB takes this idea to the next level: it was designed specifically around building custom tools for understanding specific components of the system both in situ and postmortem. On Illumos systems, the kernel ships with MDB modules that provide more than 1,000 commands to iterate and inspect various components of the kernel. Among the most frequently used is the
::stacks command, which iterates all kernel threads, optionally filters them based on the presence of a particular kernel module or function in the stack trace, and then dumps out a list of unique thread stacks sorted by frequency.
Here's an example from a system doing some light I/O:
This invocation collapsed the complexity of more than 600 threads on this system to only about seven unique thread stacks that are related to the ZFS file system. You can quickly see the state of the threads in each group (e.g., sleeping on a condition variable) and examine a representative thread for more information. Dozens of other operating-system components deliver their own MDB commands for inspecting specific component state, including the networking stack, the NFS server, DTrace, and ZFS.
Some of these higher-level analysis tools are quite sophisticated. For example, the
::typegraph command3 analyzes an entire production crash dump (without debug data) and constructs a graph of object references and their types. With this graph, users can query the type of an arbitrary memory object. This is useful for understanding memory corruption issues, where the main problem is identifying which component overwrote a particular block of memory. Knowing the type of the corrupting object narrows the investigation from the entire kernel to the component responsible for that type.
Such tools are by no means limited to production environments. On most systems, it's possible to generate a core dump from running processes too, which make core-dump analysis attractive during development as well. When testers or other engineers file bugs on application crashes, it's often easier to have them include a core dump than to try to guess what steps they took that led to the crash and then reproduce the problem from those steps. Examining the core dump is also the only way to be sure that the problem you found is the same one that the bug reporter encountered.
Higher-level dump analysis tools can be built explicitly for development as well. Libumem, a drop-in replacement for
malloc(3c) and friends, provides (among other features) an MDB module for iterating and inspecting objects related to the allocator. Combined with an optional feature to record stack traces for each allocator operation, the
::findleaks MDB command can be used to identify various types of memory leaks very quickly without having added any explicit support for this in the application itself. The
::findleaks command literally prints out a list of leaked objects and the stack trace from which each one was allocated—pointing directly to the location of each leak. Libumem is based on the kernel memory allocator, which provides many of the same facilities for the kernel.2
More importantly, dynamic languages such as Node.js are exploding in popularity as building blocks for larger distributed systems, where what might seem like a minor crash can cause cascading failures up the stack. As a result, just as with operating systems and core services, fully understanding each failure is essential to achieving the levels of reliability expected of such foundational software.
Providing a postmortem facility for dynamic environments, however, is not easy. While native programs can leverage operating-system support for core dumps, dynamic languages must present postmortem state using the same higher-level abstractions with which their developers are familiar. A postmortem environment for C programs can simply present a list of global symbols, pointers to thread stacks, and all of a process's virtual memory (all of which the operating system has to maintain anyway), but a similar facility for Java must augment (or replace) these with analogous Java abstractions. When Java programs crash, Java developers want to look at Java thread stacks, local variables, and objects, not (necessarily) the threads, variables, and raw memory used by the JVM (Java virtual machine) implementation. Also, because programs in dynamic languages run inside an interpreter or VM, when the user program "crashes," the interpreter or VM itself does not crash. For example, when a Python program uses an undefined variable (the C equivalent of a NULL pointer), the interpreter detects this condition and gracefully exits. Therefore, to support postmortem debugging, the interpreter would need to trigger the core-dump facility explicitly, not rely on the operating system to detect the crash.
Despite these difficulties, some dynamic environments do provide postmortem facilities. For example, the Oracle Java HotSpot VM supports extracting Java-level state from JVM native core dumps. When the JVM crashes, or when a core file is manually created using operating system tools such as gcore(1), you can use the jdb(1) tool to examine the state of the Java program (rather than the JVM itself) when the core file was generated. The core file can also be processed by a tool called jmap(1) to create a Java heap dump that can in turn be analyzed using several different programs.
Such facilities are only a start, however: setting up an application to trigger a core dump on crash in the first place is nontrivial. Additionally, these facilities are very specific to the HotSpot VM. There's an active Java Community Specification proposal for a common API to access debugging information, but at the time of this writing this project is stalled pending clarity about Oracle's commitment to the project.8
While the Java facility has several important limitations, many other dynamic environments do not appear to have postmortem facilities at all—at least not any that meet the constraints just described.
Python10 and Ruby4 each has a facility called a postmortem debugger, but these refer to starting a program under a debugger and having the program break into an interactive debugger session when the program crashes. This is not suitable for production for several reasons, not least of which is that it's not fully automatic. As described earlier, it's not tenable to interrupt production service while an engineer logs in to diagnose a problem interactively.
Erlang5 provides a rich crash dump facility for the Erlang runtime itself. It works much like a native crash dump in that on failure it saves a comprehensive state dump to a file and then exits, allowing the operating system to see that the program has exited and restart it immediately. The crash dump file can then be analyzed later.
The bash shell1 is interesting because its deployment model is so different even from other dynamic environments. Bash provides a mechanism called
xtrace for producing a comprehensive trace file describing nearly every expression that the shell evaluates as part of execution. This is very useful for understanding shell script failures but can produce a lot of output even for simple scripts. The output grows unboundedly as the program runs, which would normally make it untenable for production use in servers or applications, but since most bash scripts have very finite lifetimes, this mechanism is an effective postmortem facility as long as the output can be reasonably stored and managed (i.e., automatically deleted after successful runs).
uncaughtException event, which allows a program to register a function to be invoked when the program throws an exception that bubbles all the way to the top level (which would normally cause the program to crash).
JSON.stringify() and JSON.parse()).
* Synchronous functions for writing to files.
The next challenge is serializing circular objects.
JSON.stringify() does not support this for obvious reasons, so our implementation avoids this issue by pruning all circular references before serializing the debug object. While this makes it harder to find information in the dump, we know that at least one copy of every object will be present somewhere in the dump.
Given all this, the implementation is straightforward: on the
uncaughtException event, we prune circular references from the debug state, serialize it using the built-in
JSON.stringify() routine, and save the result to disk in a file called
core. To analyze the
core file, we use a tool that reads
JSON.parse() and presents the serialized state for engineers to examine. The implementation is open source and available on GitHub.7
Second, since the entire point of this system is to capture program state in the event of a crash, it must be highly reliable. This implementation is robust to most runtime failures, but it still requires additional memory first to execute the dump code and to serialize the program state. The additional memory could easily be as large as the whole heap, which makes it untenable for failures resulting from memory pressure—a common cause of failures in dynamic environments.
Third, because the implementation removes circular references before serializing the program data, the resulting dump is more difficult to browse, and the facility cannot support dumps that are not intended for postmortem analysis (such as live dumps).
Despite these deficiencies, this implementation has proved quite effective because it meets the requirements set forth earlier: it's always-on in production, fully automatic, the result is transferable to other systems for analysis, and it's comprehensive enough to solve complex problems. To address many of the scope, robustness, and richness problems described here, however, and to provide such a facility for all users of a language, the postmortem facility must be incorporated into the VM itself. Such an implementation would work similarly in principle, but it could include absolutely all program state, be made to work reliably in the face of failure of the program itself, stream the output to avoid using much additional memory, and use a format that preserves the underlying memory structures to ease understanding of the dump. Most importantly, including tools for postmortem analysis out of the box would go a long way toward the adoption of postmortem techniques in these environments.
Postmortem debugging facilities have long enabled operating-system engineers and native-application developers to understand complex software failures from the first occurrence in deployed systems. Such facilities form the backbone of the support process for enterprise systems and are essential for software components at the core of a complex software environment. Even simple platforms for recording postmortem state enable engineers to develop sophisticated analysis tools that help them quickly to root-cause many types of problems.
Meanwhile, modern dynamic languages are growing in popularity because they so effectively facilitate rapid development. Environments such as Node.js also promote programming models that scale well, particularly in the face of latency bubbles. This is becoming increasingly important in today's realtime systems.
Postmortem debugging for dynamic environments is still in its infancy. Most such environments, even those considered mature, do not provide any facility for recording postmortem state, let alone tools for higher-level analysis of such failures. Those tools that do exist are not first-class tools in their respective environments and so are not widely used. As dynamic languages grow in popularity for building critical software components, this gap is becoming increasingly important. Languages that ignore the problems associated with debugging production systems will increasingly be relegated to solving simpler, well-confined, well-understood problems, while those that provide rich tools for understanding failure postmortem will form the basis of the next generation of software bedrock.
Many thanks to Bryan Cantrill, Peter Memishian, and Theo Schlossnagle for reviewing earlier drafts of this article and to Adam Cath, Ryan Dahl, Robert Mustacchi, and many others for helpful discussions on this topic.
1. Bash Reference Manual. 2009. https://www.gnu.org/s/bash/manual/bash.html.
2. Bonwick, J. 1994. The slab allocator: an object-caching kernel memory allocator. Usenix Summer 1994 Technical Conference.
3. Cantrill, B. M. 2003. Postmortem object type identification. Fifth International Workshop on Automated and Algorithmic Debugging.
4. Debugging with ruby-debug. 2011; http://bashdb.sourceforge.net/ruby-debug.html#Post_002dMortem-Debugging.
5. Erlang Runtime System Application User's Guide, version 5.8.4. 2011. How to interpret the Erlang crash dumps; http://www.erlang.org/doc/apps/erts/crash_dump.html.
6. Gill, S. 1951. The diagnosis of mistakes in programmes on the EDSAC. Proceedings of the Royal Society A 206: 538-554.
7. joyent GitHub Project. 2011. https://github.com/joyent/node
8. Incubator Wiki. 2011. March 2011 Board reports. Kato Project; http://wiki.apache.org/incubator/March2011.
9. McGregor, D. R., Malone, J. R. 1980. Stabdump - a dump interpreter program to assist debugging. Software Practice and Experience 10(4): 329-332.
10. Python Standard Library. 2011. Python v2.7.2 documentation. pdb - the Python debugger; http://docs.python.org/library/pdb.html.
11. Schrock, E. 2009. Debugging AJAX in production. ACM Queue 7 (1); http://queue.acm.org/detail.cfm?id=1515745.
LOVE IT, HATE IT? LET US KNOW
David Pacheco is an engineer at Joyent, where he leads the design and implementation of Cloud Analytics, a realtime Node.js/DTrace-based system for visualizing server and application performance in the cloud. Previously a member of the Sun Microsystems Fishworks team, he worked on several features of the Sun Storage 7000 appliances including ZFS-based remote replication, fault management, flash device support, and HTTP/WebDAV analytics.
© 2011 ACM 1542-7730/11/0900 $10.00
Originally published in Queue vol. 9, no. 10—
see this item in the ACM Digital Library
Carlos Baquero, Nuno Preguiça - Why Logical Clocks are Easy
Sometimes all you need is the right language.
Erik Meijer, Kevin Millikin, Gilad Bracha - Spicing Up Dart with Side Effects
A set of extensions to the Dart programming language, designed to support asynchrony and generator functions
Dave Long - META II: Digital Vellum in the Digital Scriptorium
Revisiting Schorre's 1962 compiler-compiler
Bo Joel Svensson, Mary Sheeran, Ryan Newton - Design Exploration through Code-generating DSLs
High-level DSLs for low-level programming