recon

Recon wants to be a set of tools usable in production to diagnose Erlang problems or inspect production environment safely.

To build the library:

rebar3 compile

Documentation for the library can be obtained at http://ferd.github.io/recon/

It is recommended that you use tags if you do not want bleeding edge and development content for this library.

Current Status

Versions supported: OTP-17 and up. Support of R16B03-1 down to R15B02 is best effort. Builds with Rebar3 require OTP-17.1 and up because that's what the tool supports.

Changelog

Branches are organized by version. master contains the bleeding edge, 2.x contains all stable changes up to the latest release of v2, and 1.x contains all stable changes of the first version of Recon.

2.x

• 2.5.1
  • Fix support for extra messages in traces (thanks to Péter Gömöri)
  • Fix some typespecs for match specs (thanks to @chenduo)
  • Support OTP-23 change of format in allocator blocks related to carrier migration support
• 2.5.0
  • Optional formatting of records in traces (thanks to @bartekgorny)
  • Basic support for OTP-22 in recon_alloc (base handling of foreign_blocks type)
• 2.4.0
  • Optional formatting of records in traces (thanks to @bartekgorny)
• 2.3.6
  • Adapting for OTP-21. Includes the 'deprecation' of recon:files/0 since OTP-21 no longer supports listing all file descriptors, and removing error_logger_queue_len from node stats since a new logging mechanism was introduced in-process instead.
• 2.3.5
  • fixing timefold's first iteration to prevent errors at call-site by sleeping before sampling
• 2.3.4
  • fixing edoc tag that broke some downstream packaging attempts
• 2.3.3
  • fixing bin_leak arith errors
  • fixes to recon_alloc:allocators/1 (incl. R16 compatibility)
  • fix errors in scheduler wall time calculations
  • term_to_pid supports binaries
• 2.3.2
  • Allow the return_to option in recon_trace
  • More efficient sorting function for procs and ports attributes (thanks to @zhongwencool and @pichi)
  • Allow the usage of return_trace in recon_trace:calls/2-3 instead of fun(_) -> return_trace() end.
• 2.3.1
  • Updated app_deps script to run with rebar3 dependencies
  • Minor docs update
• 2.3.0
  • Doc made clearer around semantics of recon:proc_count and recon:proc_window.
  • Fix doc typos
  • Fix potential race condition on waiting for death of tracing process
  • Add an option which allows sending tracing output somewhere other than group_leader() (thanks @djnym)
  • Add ability to pass custom formatter function when tracing (thanks @iilyak)
• 2.2.1
  • Fixing type specs for recon:port_types/0 and recon_lib:count/1, thanks to @lucafavatella
  • Minor documentation fixes.
• 2.2.0:
  • Adding scheduler info metrics to get a more accurate picture than what top and CPU gives.
  • Broadening recon_trace:calls/2 interface to allow multiple match specs, which was currently only allowed for calls/3.
  • Support for mbcs_pool data in erts_alloc, and some internal refactoring, thanks to Lukas Larsson.
• 2.1.2:
  • Fixing tests for R15B02 and up
  • Fixing a backwards compatibility for R15B03 on recon_alloc operations with dumps on disk
• 2.1.1:
  • Renaming recon_trace:mfa() type to recon_trace:tspec() to avoid issues in older Erlang versions regarding redefining an existing type (Thanks Roberto Aloi)
• 2.1.0:
  • Adding recon_trace script to allow safe tracing of function calls on production nodes.
  • Adding queue_fun.awk script to inspect running functions of processes with large mailboxes in a crash dump.
• 2.0.2:
  • Preventing crashes in recon_alloc when certain expected allocators do not return results (Thanks to Michal Ptaszek)
• 2.0.1:
  • Add support for R16B03 in recon_alloc.
• 2.0.0:
  • Test suite added
  • Major rewrite of recon_alloc, thanks to Lukas Larsson. Things that changed include:
    • average_sizes/0 is renamed average_block_sizes/1 and now takes the keywords current and max.
    • Documentation updates.
    • memory/1 has new options in allocated_types and allocated_instances.
    • memory/2 has been added, which allows to choose between current and max values. memory(Term) is made equivalent to memory(Term, current).
    • Allow sbcs_to_mbcs/0 to take the arguments current and max.
    • Added unit conversion function set_unit/1, which allows to get the recon_alloc results in bytes (default), kilobytes, megabytes, and gigabytes, to help with readability.
  • Updated the internal rebar version, if anybody were to use it.
  • recon:port_info/1 no longer includes the parallelism option by default within the meta category as this would hurt backwards compatibility on some installations.
  • recon:get_state/2 is added in order to specify timeouts. recon:get_state/1 keeps its 5000 milliseconds timeout.
  • Addition of a fake attribute called binary_memory, which is callable in recon:info/2,4, recon:proc_count/2, and recon:proc_window/3. This attribute allows to fetch the amount of memory used by refc binaries for a process, and to sort by that value for counts and windows.

1.x

• 1.2.0:
  • add recon_alloc:snapshot* functions, which allow memory allocation snapshots to be taken, saved on disk, reloaded, and analyzed on-demand. Thanks to Lukas Larsson for this functionality.
  • remove parallelism data from port_info for better OTP backwards compatibility with little loss of information.
• 1.1.0:
  • add recon_lib:term_to_port to convert a string back to a usable port.
  • add recon:port_info/1 and recon:port_info/2
  • add recon_alloc module
• 1.0.0: add info/2 and info/4. The memory info type thus gets renamed to memory_used, in order to avoid conflicts when picking between a type and a specific process attribute. Types exported by the module also get updated.
• 0.4.2: extended app_deps.erl to read apps/ directories for releases
• 0.4.1: fixed bug where nodes with lots of processes could see the GC call fail if said processes failed between long calls within the bin_leak function call.
• 0.4.0: fixed bug where nodes with lots of processes or ports could see their count or window functions fail because a process or socket closed between the time the function started and before it finished. This ends up changing the API in recon_lib for the window and count functions that take a specific pid as an argument.
• 0.3.1: factored out some logic from recon:info/1 into recon_lib:term_to_pid and allowed arbitrary terms to be used for pids in recon:get_state/1.