Skip to content

ChronoLog 3.5.0

Latest

Choose a tag to compare

@github-actions github-actions released this 12 Nov 23:00
· 3 commits to main since this release

Highlights

A new dependency stack (#739). ChronoLog had built with Spack v0.21.2 since 2024, on Mercury 2.2.0, Margo 0.13.1, thallium 0.10.1 and libfabric 1.16.1. 3.5.0 builds with Spack v1.2.2 and the spack-packages recipes pinned to a commit of releases/v2026.10: Mercury 2.4.1, Margo 0.24.4, thallium 0.17.0 and libfabric 2.6.0, along with argobots 1.2, HDF5 2.2.0, Boost 1.90.0, spdlog 1.17.0, CMake 4.4.4 and pybind11 3.0.4. Margo 0.24.4, which no spack-packages release offers yet, comes from ChronoLog's own recipe in tools/spack/. The new Margo frees what Margo 0.13.1 leaked for every provider deleted explicitly, and it can bound bulk transfers and responses by a deadline, which the next release uses for the RPCs that still wait without one (#738).

  • 3.5.0 cannot talk to 3.4.0 or older. Mercury 2.2.0 sends RPC requests tagged and Mercury 2.4.1 receives them untagged, so a request between a 3.5.0 component and an older one gets no answer and no error, in either direction. Upgrade the visor, keepers, graphers, players and every client together.
  • A replay that starts at time 0 is answered (#741). When the player planned an archive read for such a replay, it dropped the request, and the client got CL_ERR_QUERY_TIMED_OUT after 180 s. The player now starts the replay at 1, and an archive request the player refuses still answers its query, short, so the client gets CL_ERR_PARTIAL_RESULT.
  • Stories start after a keeper left (#743, #707). Once one keeper of a recording group stopped, every new AcquireStory returned CL_ERR_NO_KEEPERS although the other keepers were running: deleting the departed keeper's admin client in the visor took away the RPCs that the other admin clients used. Per-peer RPC clients no longer deregister the RPCs they share.
  • Docs and tests. The 3.5.0 docs describe the Spack v1.2.2 setup, warn on the client pages about mixed versions, and say that 3.4.0 and older build with Spack v0.21.2; the native and Docker tutorials use chrono-client-cli and chrono-bench, and the chrono-mcp setup guide follows the current install. The Docker images build on Spack v1.2.2. The unit suite grows from 758 to 764 tests, and the durable-send tests that kill a keeper run it in a child process.

Upgrading from 3.4

  • Build from source with Spack v1.2.2 and gfortran. Spack needs a Fortran compiler to build a pybind11 build dependency. Spack v0.21.2 cannot read the new spack.yaml, and v1.2.2 cannot build the spack.yaml of 3.4.0 and older, so keep Spack v0.21.2 for those versions. In an existing checkout, the deploy scripts move the recipe checkout to the pinned commit, and a spack.lock from Spack v0.21.2 is concretized again on its own. See Environment Setup.
  • Clients. The protocol version stays 3, but an older client's Connect() waits without an answer instead of returning CL_ERR_PROTOCOL_VERSION_MISMATCH; rebuild C++ clients and reinstall the Python module.
  • Python. py_chronolog_client is built for the Python of ChronoLog's Spack environment, now Python 3.14 (3.4.0: 3.11), and no other Python imports it. Run Python clients with that interpreter and PYTHONPATH on the installation's lib directory; no LD_LIBRARY_PATH or symbolic link is needed. See the Python client.
  • Configuration and the archive. No configuration key and no client API change. Archives written by 3.4.0 stay readable: a fresh 3.5.0 deployment replayed every event of a story that a stopped 3.4.0 deployment had archived in the same directory.
  • Stop 3.4.0 cleanly first. --stop reports durable events still pending in a keeper WAL; let them reach the archive before starting 3.5.0.

Behavior changes

  • chrono-bench reports its Record-event and Replay-story throughputs in events per second, as labelled; they were in millions of events per second.
  • chrono-bench -r reports an unexpected ReplayStory code and goes on, instead of aborting.
  • docker/dynamic_deploy.sh deploys again: it looked for tools/deploy_cluster.sh, a path no release since 3.0.0 installs.

Known issues

  • Most RPCs still have no deadline (#738). Over ofi+tcp, an accepted write to a dead keeper, and an acquire while the visor calls that keeper, wait until the transport gives up; in keeper_restart_held_test.sh the acquire did not finish within 15 s. The shipped configuration uses ofi+sockets.
  • deploy_local.sh returns before the keepers have registered with the visor, so a client started at once can get CL_ERR_NO_KEEPERS, and the basic, writer and telemetry writer examples then crash on the null story handle: #742.
  • A failed story start leaves the story bound to its group, and its events are lost: #725.
  • Two reader-mode clients on one host both bind port 5557: #719.
  • After a keeper restart, tail reads (playback) do not return the recovered events; a replay does.
  • Other open items: #708, #709 (failed archive writes), #711, #712, #715 (resource use), #716, #717, #718 (deployment).

Full Changelog: v3.4.0...v3.5.0