mars/ChangeLog
2016-03-09 09:51:28 +01:00

637 lines
27 KiB
Plaintext

IMPORTANT: the historic distinction between MARS Light and the future
MARS Full has been dropped. Now all versions are simply called "mars".
Old tagnames light* will remain valid, but newer names will follow the
convention s/light/mars/g (this means that the old version number counting
will be continued, only the "light" is substituted).
Meaning of stable tagnames
--------------------------
Example: mars0.1stable01:
0 = version of on-disk data structures
(only incremented when downgrades are impossible)
(not incremented on backwards-compatible upgrades)
1 = version of feature set
stable = feature set is frozen during this series
01 = bugfix revision
Example: mars0.2beta2.3:
The general idea is as before.
"beta" means that new features are roughly tested
in the lab, but not in production, so there may be
some bugs. New features may be added during
the beta phase.
Example: mars0.3alpha*:
Never use this for production. Only for historic
code inspection.
Release Conventions / Branches / Tagnames
-----------------------------------------
mars0.1 series (stable):
- Asynchronous replication.
Currently operational at more than 1700 servers at
1&1, more than 9,000,000 operating hours (Mar 2016)
- Unstable tagnames: light0.1beta%d.%d (obsolete)
- Stable branch: mars0.1.y
- Stable tagnames: mars0.1stable%02d
mars0.2 series (currently in beta stage):
Mostly for internal needs of 1&1 (but not limited to that).
- Getting rid of the kernel prepatch! MARS may be built
as an external kernel module for any supported
kernel version. First prototype is only tested for
unaltered 3.2.x vanilla kernel, but compatibility to
further vanilla kernel versions (maybe even
Redhat-specific ones) will follow during the course of
the MARS mars0.2 stable series. The problem is not
compatibility as such, but _testing_ that it really
works. These tests need a lot of time.
=> further arguments for getting to kernel upstream ASAP.
- Improved network throughput by parallel TCP connections
(in particular under packet loss).
Also called "socket bundling".
First benchmarks show an impressive speedup over
highly congested long-distance lines.
- Future-proof updates in the network protocol:
Mixed operation of 32/64bit and/or {big,low}endian
- Support for multi-homed network interfaces.
- Transparent data compression over low bandwidth lines.
Consumes a lot of CPU, therefore only recommended for
low write loads or for desperate network situations.
- Remote device: bypassing iSCSI. In essence,
/dev/mars/mydata can appear at any other cluster member
which doesn't necessarily need any local disks.
- Various smaller features and improvements.
- Unstable tagnames: mars0.2beta%d.%d (current)
- Stable branch: mars0.2.y (already in use for beta)
- Stable tagnames: mars0.2stable%02d (planned)
mars0.3 series (planned):
(some might possibly go to 1.0 series instead)
- Improve replication latency.
- New pseudo-synchronous replication modes.
For the internal needs of database folks at 1&1.
- (Maybe) old test suite could be retired, a new
one is at github.com/schoebel/test-suite
- Unstable tagnames: mars0.3beta%d.%d (planned)
- Stable branch: mars0.3.y (planned)
- Stable tagnames: mars0.3stable%02d (planned)
mars1.0 series (planned):
- Replace symlink tree by transactional status files
(future-proof)
This is required for upstream merging to the kernel.
It has further advantages, such as better scalability.
- Trying to additionally address public needs.
- Potentially for Linux kernel upstream,
- Unstable tagnames: mars1.0beta%d.%d (planned)
- Stable branch: mars1.0.y (planned)
- Stable tagnames: mars1.0stable%02d (planned)
full* (somewhen in future)
WIP-* branches are for development and may be rebased onto anything
at any time without notice. They will disappear eventually.
*stable* branches mean that only bug fixes and documentation
updates / clarifications will be applied. Updates to the test suite /
new test cases potentially disguising bugs, and other minor additions
of debugging code / paranoia code which may lead to discovery
of bugs are also possible. Error messages / warnings and their
error class may also be changed.
NO NEW FEATURES, not even minor ones, except when absolutely
necessary for a bugfix, or for an important usability improvement
(such as clearer display of errors, hints for resolving them, etc).
-----------------------------------
Changelog for series 0.2:
(you need to checkout branch mars0.2.y to see any details)
-----------------------------------
Changelog for series 0.1:
mars0.1stable29
--------
* Minor fix: on very fast hardware and networks, sync could take
a while for terminating.
* Minor fix: external module build.
* Major usability improvement: new expert commands marsadm
lowlevel-ls-host-ips, lowlevel-set-host-ip, lowlevel-delete-host.
Necessary for moves between networks, dedicated replication IPs,
etc.
* Minor doc update.
mars0.1stable28
--------
* Doc: describe new naming conventions.
MARS Light is now simply called MARS.
No distinction between "Light" and the future "Full" anymore.
Please note that the git branches light0.1.y and light0.2.y have
been renamed to mars0.1.y and mars0.2.y respectively.
* Minor sourcecode cleanup: s/light//g or s/light/main/g
where appropriate.
No other changes in the sourcecode, deliberately.
In case anyone encounters any build problems compiling MARS,
this release is separated just for the sake of build testing,
or Debian packaging testing, etc.
* Doc: minor clarifications.
mars0.1stable27
light0.1stable27
--------
* Critical fix: typo in sync progress comparison code could lead
to data version mismatches during sync when alternating with
replay. Only observed at a certain new hardware class, and only
while testing with an extremely high load (9 loaded resources
in parallel to 9 concurrent syncs). As a workaround,
echo 0 > /proc/sys/mars/sync_flip_interval_sec can be used.
Nevertheless, update is highly recommended!
* Major fix: slow memory leak (regression from light0.1stable26).
Only when starting the transaction logger (i.e. primary is typically
not affected). But don't let run it for a longer time.
Monitoring is possible via /proc/slabinfo (size-64 or siblings).
* Minor fix: join-cluster did not check for duplicate IP addresses.
* Minor fixes: some unnecessary annoying error messages.
* Docu: new slides from GUUG 2016 in Köln.
light0.1stable26
--------
* Minor fixes: some primitive macros were reporting misleading or
even wrong values at split brain, or during/after emergency mode.
Some high-level macros as well as try_to_avoid_split_brain
should work better / more reliable now.
* Minor fix: potential deadlock after crash reboot, or after
defective /mars filesystem. Never observed in practice.
* Minor safeguard: unnecessary split brain could emerge at
secondaries under extremely rare and strange conditions.
Unsure whether it ever occurred in practice.
* Minor usability improvement: show incorrect permissions on /mars.
Some other sysadmin tools like Puppet seem to have their own
default notion of "secure permissions" ;)
* Minor doc reorg, better chapter structure.
light0.1stable25
--------
* Major fix: in rare cases "marsadm primary" (without --force)
could go into an endless loop, even if --timeout= was specified.
* Minor fix: in rare cases of hanging or defective IO, crashes
of the primary could replicate versionlinks to the secondary,
but after reboot they were missing at the primary because of
of hanging IO or other IO / RAID controller problems.
Now using sync_filesystem() for either ensuring actuality,
or for letting the mars_light main control thread hang
(which will hopefully be noticed soon by monitoring).
* Minor fix: join-cluster uses rsync, which could abort due to
vanished filesystem objects while the primary is actively running.
Now it should tolerate such "errors".
* Minor fixes / additions at primitive macros.
* Tiny doc update.
light0.1stable24
--------
* Skip this release due to a regression.
light0.1stable23
--------
* Minor fix: the new replay-code error message was forgotten
to reset at secondaries. Now the annoying old error message
disappears after the next successful logrotate.
* Minor fixes of internal marsadm code (not in use until now).
* Minor doc update.
light0.1stable22
--------
* Critical fix for non-storage servers: the /mars directory
was readable by ordinary non-root users, opening a potential
security hole. Originally MARS was designed for standalone
storage servers solely, but now it is increasingly deployed to
machines where ordinary users can log in.
Update recommended, but only urgent for potentially affected
installations.
* Minor fix: when a logfile was damaged (observed at defective
hardware), this was often (but not always) detected by the
md5 data checksums in the transaction logfiles. So far so good.
The replay / recovery process stopped for a very good reason.
But it was not easily possible to _force_ any of the resource
members into primary role when the defect was already present at
the _primary_ (which happend once during 7 millions of operating
hours, and at a primary site which proved defective afterwards),
and the defect had been replicated to all secondaries.
As a workaround, the resource could be destroyed via leave-resource
everywhere, and re-surrected from scratch. Clumsy.
Now an md5 checksum error in the middle of a logfile is
treated similarly to an EOF. "primary --force" will succeed now,
without applying the defective data (as before).
Split brain will result for sure in such a case.
* Minor improvement: md5 logfile checksum errors are now displayed
directly in the diskstate macro (and therefore also at plain
"view").
* Minor improvement: when "marsadm view all" told you "InConsistent"
as the disk state, this was _formally correct_ because it related
to the state of the _disk_, not to the state of the replication.
The former message could appear regularly during ordinary
out-of-order writeback at the primary side, without violating
the consistency of /dev/mars/mydata.
However, many people were confused and alarmed by the irritating
message.
Now a better wording is used: "WriteBack" and "Recovery" describes
more intuitively what is really happening :)
* Minor doc improvements.
light0.1stable21
--------
* Hint: now MARS has been rolled out to more than 1600 servers,
including some MySQL database servers, and has collected more
than 6 millions of operation hours.
* Minor fixes, none of them observed in practice, only found
by testing while working on new features:
- potential read page fault
- potential deadlock
- incorrect remote symlink update under untypical circumstances
light0.1stable20
--------
* Hint: MARS is now running on more than 850 storage servers,
and has collected more than 4.5 millions of operation hours.
There were no new incidents with customer impact since the last
major bugfix (more than 3 millions of operation hours since then).
It is difficult to deduce a reliability from that, but it appears
that at least 99.999%, if not 99.9999% are now real for the
MARS component as a standalone component (not to be confused with
overall system reliability). Our storage hardware is clearly much
less reliable. MARS does compensate these defects all the time.
* Minor fix: memory leak in networking code, does not occur
at light0.1 operations (but maybe future versions of MARS).
* Doc: add presentation slides from Froscon2015.
light0.1stable19
--------
* Minor safeguard: warn when somebody tries leave-resource --host=
for a damaged host, and later the dead host resurrects in an
unreasonable way.
* Doc update: describe use cases for DRBD vs MARS more clearly.
* Minor spelling fixes.
light0.1stable18
--------
* Minor safeguard: prevent join-resource when previous log-purge-all
has been forgotten. Prevent create-resource also when previous
delete-resource has been forgotten. Anyway, this happens only in
very exotic repair scenarios after very heavy failures.
* Doc updates: simplify descriptions of split-brain resolution and
emergency mode resolution. Nowadays 'invalidate' will do everything
in all tested cases; the more complex alternative methods have
been moved to the appendix.
light0.1stable17
--------
* Minor fix: stacktrace / oops in aio callback path due to a
subtle race, observed once during 2.5 millions of operation hours.
In the observed case, the secondary was hanging, without
customer impact. However, the error class could potentially
occur also at the primary side. Probably the bug was triggered
by a hardware problem from the RAID controller.
light0.1stable16
--------
* Minor fix: sync could take a long time to complete under high
application load, similarly to a live-lock.
* Some smaller minor fixes for annoying messages.
* Contrib: added configurable Nagios check.
* Contrib: added some example scripts which could be used by
clustermanagers etc.
* Doc: important new section on pitfalls when using existing
clustermanagers UNMODIFIED for long distance replication.
PLEASE READ!
light0.1stable15
--------
* NOTICE: MARS succeeded baptism on fire at 04/22/2015 when a whole
co-location had a partial power blackout, followed by breakdown
of air conditioning, followed by mass hardware defects due to
overheating. MARS showed exactly 0 errors when (emergency)
switching to another datacenter was started in masses.
* Major fix of race in transaction logger: the primary could hang
when using very fast hardware, typically after ~24000 operation
hours. The problem was noticed 6 times during a grand total of
more than 1,000,000 operation hours on a mixed hardware park,
showing up only on specific hardware classes. Together with 3
other incidents during early beta phase which also had customer
impact, this means that we have reached a reliability of about
===> 99.999%
After this fix, the reliability should grow even higher.
A workaround for this bug exists:
# echo 2 > /proc/sys/mars/logger_completion_semantics
Update is only mandatory when you cannot use the workaround.
* Minor improvement in marsadm: re-allow --force combined with "all".
This is highly appreciated for speeding up operations / handling
during emergency datacenter switchover.
* Various smaller improvements.
* Contrib (unsupported): example rollout script for mass rollout.
light0.1stable14
--------
* Minor safeguard: modprobe mars will refuse to start when the
cluster UUID is missing.
* Minor fix: external race in marsadm resize, only relevant
for scripting.
* Minor fix: potential race on plugged IO requests.
* Clarify output of marsadm view. Many systematical improvements
and hints.
* Add some unevitable macros for scripting / automation.
* Various tiny improvements.
light0.1stable13
--------
* Critical safeguard for accidental join-cluster with wrong argument:
make UUID mandatory, disallow completely unrelated hosts to
communicate symlink tree updates when their UUIDs mismatch.
* Minor fix: leave-resource --host=other did not work when disks
were named differently throughout the cluster.
* Minor fix: detach --host=other --force (which is needed as a
precondition) did not work.
* Various minor fixes and clarifications. "marsadm view all"
now reports the communication status in the cluster.
light0.1stable12
--------
* Critical (but usually not extremely relevant) fix:
When emergency mode occurs just during a sync, the target could
remain inconsistent without notice. Now noticed.
You always could/should manually invalidate whenever an
emergency mode appeared.
Now this is automatically fixed by restarting any sync from
scratch (if one was actually running before; otherwise consistency
was never violated).
* Major documentation update / corrections.
* Major (but less relevant) fix: leave-cluster did not really work.
* Minor fix (regression): rmmod could hang when sync was running.
* Various minor fixes and clarifications.
light0.1stable11
--------
* Major documentation update. mars-manual.pdf increased from
66 to 80 pages. Please read! You probably should know this.
* Minor fixes: better cleanup on invalidate / leave-resource.
* Minor clarifications: more precise EIO error codes, more verbose
error reporting via "marsadm cat".
light0.1stable10
--------
* Major fixes of internal network protocol errors, leading to
internal shutdown of sockets, which were transparently re-opened.
It could affect network performance. Not sure whether
stability was also affected (probably under extremely high load);
for better safety you should upgrade.
* Major fix from Manuel Lausch: regex parsing sometimes went
completely wrong when hostnames followed a similar name scheme
than internal symlinks.
* Major, only relevant for k>2 replicas: fix wrong internal sharing
of data structures resulting from parallel data connections.
* Minor fix: race in fake-sync.
* Minor fix: race in invalidate.
* Minor, only for k>2 replicas: fix direct primary handover when
some non-involved hosts are currently unreachable.
* Minor: improve becoming primary during split brain.
* Minor: improve becoming primary when emergency mode starts.
* Minor: silence some annoying stderr messages.
* Several internal minor fixes and clarifications.
light0.1stable09
--------
* Major fix of scarce race (potentially critical): the bio response
thread could terminate too early, leading to a premature dealloc
of kernel memory. This has only been observed on slow virtual
machines with slow virtual devices, and very high load on k=4
replicas. This could potentially affect the stability of the system.
Although not observed at production machines at 1&1, I recommend
updating production machines to this release ASAP.
* Major usability fix: incorrect commandline options of marsadm
were just ignored if they appeared after the resource argument.
Misspellings could cause undesired effects. For instance,
"marsadm delete-resource vital --force --MISSPELLhost=banana"
was accidentally destroying the primary during operation (which
is _possible_ when using --force, and this was even a _required_
sort of "STONITH"-like feature -- however from a human point
of view it was intended to destroy _another_ host, so this was
an unexpected behaviour from a sysadmin point of view).
* Major workaround: the concept "actual primary" is wrong, because
during split brain there may exist several primaries. Do not
use the macro view-actual-primary any longer. It is deprecated now.
Use view-is-primary instead, on each host you are interested in.
* Minor fix: "marsadm invalidate" did not work in some weired
split brain situations / was not equivalent to
"marsadm leave-resource $res; marsadm join-resource $res".
The latter was the old workaround to fix the situation.
Now it shouldn't be necessary anymore.
* Minor fix: pause-fetch could take very long to terminate.
* Minor fix: marsadm wait-cluster did not wait for all hosts
particiapting in the resource, but only for one of them.
This is only relevant for k>2 replicas.
* Minor fix: the rates displayed by "marsadm view" did not drop down
to 0 when no progress was made.
* Minor fix: logging to syslog was incomplete.
* Minor usability fix: decrease boring speakyness of "log-rotate"
and "log-delete" for cron jobs.
* Minor fixes: several internal awkwardnesses, potentially affecting
performance and/or stability in weired situations.
light0.1stable08
--------
* Minor fix: after emergency mode, a versionlink was forgotten
to create. This could lead to unnecessary reports of split
brain and/or need for additional re-invalidate.
* Minor fix: the predicate 'view-is-consistent' reported 'false'
in some situations on secondaries when all was ok.
* Minor fix: it was impossible to determine the 'is-consistent'
from 'marsadm view' (without -1and1 suffix). Added a new [Cc-]
flag. This is absolutely needed to determine whether the
underlying disks must have the same checksum (provided that
both disks are detached and the network works and fetch+replay
had completed before the detach).
* Updated docs to reflect this.
* Minor fix: 'invalidate' did not work when the resource was not
completely detached. Now it implicitly does a detach before
starting invalidation.
* Minor fix: wait-umount was waiting for umount of _all_ primaries
during split brain. Now it waits only for umount of the local node.
Notice that having multiple primaries in parallel is an
erroneous state anyway.
* Minor fix: leave-cluster did not work without --force.
light0.1stable07
--------
* Minor fix: re-creation of a completely destroyed resource
did not always work correctly
light0.1stable06
--------
* Major fix: becoming primary was hanging in scarce situations.
* Minor fix: some split brains were not always detected correctly.
* Minor fix for Redhat openvz kernel builds.
* Several fixes for 1&1 internal Debian builds.
light0.1stable05
--------
* Major fix: incomplete calls to vfs_readdir()
which could lead to incomplete symlink updates /
replication hangs.
* Minor fix: scarce race on replay EOF.
* Separated kernel from userspace build environment.
* Removed some potentially dangerous Kconfig options
if they would be set to wrong values (robustness against
accidentally producing bad kernel modules).
* Dito: some additional checks against bad main Kconfig options
(mainly for out-of-tree builds).
* Separated contrib code from maintained code.
* Added some pre-patches for newer kernels
(WIP - not yet fully tested at all combinations)
* Minor doc addition: LinuxTag 2014 presentation.
light0.1stable04
--------
* Quiet annoying error message.
* Minor readability improvements.
* Minor doc updates.
light0.1stable03
--------
* Major: fix internal aio race (could lead to memory corruption).
* Fix refcounting in trans_logger.
* Some minor fixes in module code.
* Fix 1&1-internal out-of-tree builds.
* Various minor fixes.
* Update monitoring tools / docs (German, contributed by Jörg Mann).
light0.1stable02
--------
* Fix sorting of internal data structure.
* Fix IO error propagation at replay.
light0.1stable01
--------
* Fix parallelism of logfile propagation: sometimes a secondary
could get a more recent version than the primary had on stable
storage after its crash, eventually leading to an (annoying)
split brain. Some people might take this as a feature instead
of a bug, but now the logfile transfer starts only after the
primary _knows_ that the data is successfully committed to
stable storage.
* Fix memory leaks in error path.
* Fix error propagation between client and server.
* Make string allocation fully dynamic (remove limitation).
* Fix some annoying messages.
* Fix usage output of marsadm.
* Userspace: contributed bugfix for Debian udev rules by Jörg Mann.
* Improved debugging (only for testing).
light0.1beta0.18 (feature release)
--------
* New commands marsadm view-$macroname
* New customizable macro processor
* New err/warn/inf reporting via symlinks
* Per-resource emergency mode
* Allow limiting the sync parallelism
* New flood-protected syslogging
* Some smaller improvements
* Update docs
* Update test suite
light0.1beta0.17
--------
* Major bugfix: race in logfile switchover could sometimes
lead to the wrong logfile (extremely rare to hit, but
potentially harmful).
* Disallow primary switching when some secondaries are
syncing.
* Fix logfile fetch from multiple peers.
* Fix computation of transitive closure (affected
log-purge-all, split brain detection, and many others).
* Fix incorrect emergency mode detection.
* Primaries no longer fetch logfiles (unnecessarily, only
makes a difference at concurrent split brain operations).
* Detached resources no longer fetch logfiles (unexpectedly).
* Myriads of smaller fixes.
light0.1beta0.16
--------
* Critical bugfix: "marsadm primary --force" was assumed to be given
by sysadmins only in case of emergency, when the network is down.
When given in non-emergency cases where the old primary continues
to run (/dev/mars/* being actively used and written), the
old primary could suddendly do a "logrotate" to the
new split-brain logfile produced by the new (second) primary.
Now two primaries should be able to run concurrently in split-brain
mode without mutually trashing their logfiles.
* primary --force now only works in disconnected mode, in order
to hinder unintended forceful creation of split brain during
normal operation.
* Stop fetching of logfiles behind split brain points (save space
at the target hosts - usually the data will be discarded later).
* Fixed split brain detection in userspace.
* leave-resource now waits for local actions to take place
(remote actions stay asynchronously).
* invalidate / join-resource now work only if a designated primary
exists (otherwise they would not know uniquely from whom
to start initial sync).
* Update docs, clarify scenarios intended <-> emergengy switching.
* Fixed mutual overwrite of deletion symlinks in case of racing
log-deletes spawned in parallel by cron jobs (resilience).
* Fixed races between deletion and re-erection (e.g. fresh
join-resource after leave-resource during network partitions).
* Fixed duration of network timeouts in case the network is down
(replaced non-working TCP_KEEPALIVE by explicit timeouts).
* New option --dry-run which does not really create symlinks.
* New command "delete-resource" (VERY DANGEROUS) for
forcefully destroying a resource, even when it is in use.
Intended only for _emergency_ cases when sysadmins are
desperate. Use only by hand, first run with --dry-run in order
to check what will happen!
* New command "log-purge-all" (potentially DANGEROUS) for
resolving split brain in desperate situations (cleanup of
leftovers). Only use by hand, first run with --dry-run!
* Lots of smaller imprevements / usability / readability etc.
* Update test suite.
light0.1beta0.15
--------
* Introduce write throttling of bulk writers.
* Update test suite.
light0.1beta0.14
--------
* Fix logfile transfer in case of "holes" created by
emergency mode.
* Fix "marsadm invalidate" after emergency mode had been entered.
* Fix "marsadm resize" capacity propagation from underlying LVM.
* Update test suite.
light0.1beta0.13
--------
* Fix shutdown during operation (flying requests).
* Fix unnecessary Lamport clock propagation storms.
* Improve unnecessary page cache utilisation (mapfree).
* Update test suite.
light0.1beta0.12 and earlier
--------
There was no dedicated ChangeLog. For details, look at the
commit history.
Release Policy / Software Lifecycle
-----------------------------------
New source releases are simply announced by appearance of git tags.