rezolus

Fixed

• Fixes potential deadlock by updating dashmap dependency.
• Updates other dependencies to pull-in bugfixes.
• Improves memcache sampler reconnect logic.

Fixed

• Fixes scheduler runqueue latency BPF for newer kernels (>= 5.14) and those
  built with gcc >= 10
• Fixes issue with release 2.16.1

Added

• Adds a new process sampler which can monitor CPU and memory utilization for
  a process. (#282)

Fixed

• Fixes issue where release archives do not build successfully due to inclusion
  of vergen in the build script. (#279)

Fixed

• Fixes tcp/connection/accepted and tcp/connection/initiated metrics on
  kernel 5.10. (#266)
• Fixes tcp/receive/duplicate and tcp/receive/out_of_order metrics. (#267)

Changed

• Allow selective enablement of various BPF metrics. (#254)
• Support up to BCC 0.23.0 and makes it the new default version. (#256)
• Removed ssl support in http sampler to remove dependency on openssl. (#257)

Added

• Adds TCP jitter and connections accepted and initiated using BPF. (#247)
• Adds Pelikan specific stats to memcache sampler. (#249)
• Adds TCP packet drops counter using BPF. (#250)
• Adds TCP tail loss recovery and retransmit timeout using BPF. (#253)
• Adds TCP duplicate segment and out-of-order segment counters using BPF. (#255)

Fixed

• Improved handling of BPF initialization errors so that samplers will continue
  to initialize remaining BPF probes if fault tolerant error handling is
  enabled. (#259)

Added

• Adds new SRTT metric for TCP sampler using BPF. (#238)
• Adds new krb5kdc sampler to get telemetry on MIT Kerberos. (#241)

Fixed

• Interrupt sampler failed to sample all interrupts if it encountered an
  unexpected keyword. (#225)
• Interrupt sampler incorrectly initialized per-NUMA node counts for NVMe and
  network interrupts. (#226)
• Memory sampler failed to report some stats. (#227)
• CPU c-state sampling now handles older style c-state names. (#229)
• Prometheus metric exposition now includes type annotations and changes the
  format for percentiles to be encoded as a label value. This fixes collection
  with OpenTelemetry. (#230)

Changed

• Removed unused interrupt/serial metric from the interrupt sampler. (#228)

Fixed

• NTP sampler failed to build with musl toolchain. (#216)

Added

• New usercall sampler for probing arbitrary userspace functions in shared
  libraries.

Fixed

• HTTP and Memcache samplers reporting incorrect percentiles.

Added

• Nvidia GPU sampler which uses the Nvidia Management Library (NVML) to gather
  telemetry for GPU utilization and health.
• NTP sampler to gather telemetry about NTP synchronization.

Fixed

• Disk BPF sampling now compatible with newer kernels.
• Bug introduced in 2.8.0 caused sample rates greater than 1000ms to cause
  errors.

Changed

• Updates tokio to 0.3.1 from 0.2.x
• Reduces syscall load by reusing filehandles in memory, interupt, and network
  samplers.

Added

• Page Cache sampler which uses BPF to instrument Page Cache hit/miss.

Fixed

• Updated rustcommon dependencies to get some runtime performance benefits.
• Added proper core -> NUMA node mapping to address issues with per-node metrics
  for interrupt sampler.
• Reduce the cost of disabled samplers by skipping all initialization of
  samplers which are not enabled in the config.
• Documentation updates.

Changed

• Metrics library has been replaced with a new version which reduces memory
  footprint.
• Samplers have been optimized to reduce number of system calls and temporary
  allocations.
• Arbitrary percentiles may now be expressed in the configuration.
• Percentile exposition format has changed to allow arbitrary percentiles. They
  are now expressed in a decimal format padded to 2 digits before the decimal.
  For example, the 5th percentile is now p05 and the 99.9th percentile is now
  p99.9.

Fixed

• Fixed memcache sampler causing tokio worker to panic due to issues registering
  the tcp stream with the tokio runtime.

Changed

• Perf event sampling now implemented with BPF. Now requires building with BPF
  support.
• Renamed worker threads and set limit for total number of runtime threads
  instead of just core threads.

Added

• CPU sampler now includes CPU frequency.

Fixed

• BPF probes are now dropped properly on program termination. Previously, on
  some kernel versions, BPF probes might remain after exit.
• Memcache sampler was not being initialized. It's now re-enabled.

Added

• Expanded memory sampler coverage to include telemetry related to NUMA access
  patterns, transparent hugepages, and compaction.

Fixed

• Disk sampler was not reporting stats for all disks on some multi-disk systems.

Added

• Interrupt sampler now has BPF sampling of time distribution of hardirq/softirq
  handlers.

Fixed

• Replaced remaining uses of chashmap with dashmap which has better performance
  characteristics.
• Statically linking bcc/bpf has fixes in upstream crates.

Added

• HTTP sampler to poll JSON endpoint and provide summary metrics
• Added support for bcc 0.15.0, making it the new default version

Added

• TCP abort metrics added to tcp sampler
• Increased max for context switch histogram to prevent clipping

Fixed

• Fixed bug where percentiles could get stuck at the max value if they hit it