1 2009-03-20 Andreas J. Koenig <andk@cpan.org>
3 * bug on the mirroring slave: when the dirtymark gets increased we
4 probably do not reset the done intervals. The mirrorer stays within
5 tight bounds where it tries to sync with upstream and never seems to
6 finish. In the debugging state file I see lots of identical intervals
7 that do not get collapsed. When I restart the mirrorer it dies with:
9 Sync 1237507989 (227/33111/Z) id/X/XI/XINMING/Catalyst-Plugin-Compress.tar.gz ...
10 _bigfloatcmp called with l[1237400817.94363]r[UNDEF]: but both must be defined at /home/k/sources/rersyncrecent/lib/File/Rsync/Mirror/Recentfile/FakeBigFloat.pm line 76
11 File::Rsync::Mirror::Recentfile::FakeBigFloat::_bigfloatcmp(1237400817.94363, undef) called at /home/k/sources/rersyncrecent/lib/File/Rsync/Mirror/Recentfile/FakeBigFloat.pm line 101
12 File::Rsync::Mirror::Recentfile::FakeBigFloat::_bigfloatge(1237400817.94363, undef) called at /home/k/sources/rersyncrecent/lib/File/Rsync/Mirror/Recentfile/Done.pm line 226
13 File::Rsync::Mirror::Recentfile::Done::_register_one('File::Rsync::Mirror::Recentfile::Done=HASH(0x84c6af8)', 'HASH(0xb693f2dc)') called at /home/k/sources/rersyncrecent/lib/File/Rsync/Mirror/Recentfile/Done.pm line 200
14 File::Rsync::Mirror::Recentfile::Done::register('File::Rsync::Mirror::Recentfile::Done=HASH(0x84c6af8)', 'ARRAY(0x8c54bfc)', 'ARRAY(0xb67e618c)') called at /home/k/sources/rersyncrecent/lib/File/Rsync/Mirror/Recentfile.pm line 1044
15 File::Rsync::Mirror::Recentfile::_mirror_item('File::Rsync::Mirror::Recentfile=HASH(0x84abcb0)', 227, 'ARRAY(0x8c54bfc)', 33110, 'File::Rsync::Mirror::Recentfile::Done=HASH(0x84c6af8)', 'HASH(0x84abd8c)', 'ARRAY(0x839cb64)', 'HASH(0x839c95c)', 'HASH(0xb6a30f0c)', ...) called at /home/k/sources/rersyncrecent/lib/File/Rsync/Mirror/Recentfile.pm line 992
16 File::Rsync::Mirror::Recentfile::mirror('File::Rsync::Mirror::Recentfile=HASH(0x84abcb0)', 'piecemeal', 1, 'skip-deletes', 1) called at /home/k/sources/rersyncrecent/lib/File/Rsync/Mirror/Recent.pm line 564
17 File::Rsync::Mirror::Recent::_rmirror_mirror('File::Rsync::Mirror::Recent=HASH(0x84ab6d4)', 7, 'HASH(0x8499488)') called at /home/k/sources/rersyncrecent/lib/File/Rsync/Mirror/Recent.pm line 532
18 File::Rsync::Mirror::Recent::rmirror('File::Rsync::Mirror::Recent=HASH(0x84ab6d4)', 'skip-deletes', 1) called at /home/k/sources/CPAN/GIT/trunk/bin/testing-rmirror.pl line 27
20 and debugging stands at
26 epoch: 1237400802.5789
27 path: id/J/JO/JOHND/CHECKSUMS
30 epoch: 1237400807.97514
31 path: id/M/MI/MIYAGAWA/CHECKSUMS
34 epoch: 1237400817.94363
35 path: id/X/XI/XINMING/Catalyst-Plugin-Compress.tar.gz
43 and it is reproducable.
45 Why does the mirrorer not fetch a newer Z file? It is 18 hours old while
46 pause has a fresh one.
48 2009-03-17 Andreas J. Koenig <andk@cpan.org>
50 * done: verified the existence of the floating point bug in bleadperl
51 and verified that switching from YAML::Syck to YAML::XS does not resolve
54 BTW, the switch was doable with
56 perl -i~ -pe 's/Syck/XS/g' lib/**/*.pm t/*.t
58 and should be considered as a separate TODO
60 * todo: integrate a dirty update with two aggregate calls before
61 unlocking for frictionless dirtying
63 * todo: start the second rsync daemon on pause
65 * todo: move index files to .recent: this cannot simply be done by
66 setting filenameroot to .recent/RECENT. Other parts of the modules rely
67 on the fact that dirname(recentfile) is the root of the mirrored tree.
69 2009-03-16 Andreas J. Koenig <andk@cpan.org>
71 * What was the resolution of the mirror.pl delete hook bug? Do we call
72 the delete hook when pause removes a file from MUIR?
74 * Today on pause: Updating 2a13fba..29f284d and installing it for
75 /usr/local/perl-5.10.0{,-RC2}
77 TURUGINA/Set-Intersection-0.01.tar.gz was the last upload before this
78 action and G/GW/GWILLIAMS/RDF-Query-2.100_01.tar.gz the first after it
80 2009-03-15 Andreas J. Koenig <andk@cpan.org>
82 * currently recent_events has the side effect of setting dirtymark
83 because it forces a read on the file. That should be transparent, so
84 that the dirtymark call always forces a cache-able(?) read.
86 * The bug below is -- after a lot of trying -- not reproducible on a
87 small script, only in the large test script. The closest to the output
93 my $x = "01237123229.8814";
98 ($l,$r) = ($1,$2) if $x =~ /(.)(.+)/;
100 $l = "1237123231.22458";
101 $r = "1237123231.22458";
103 Devel::Peek::Dump $l;
104 Devel::Peek::Dump $r;
105 Devel::Peek::Dump $x = $l <=> $r;
109 The checked in state at c404a85 fails the test with my
110 /usr/local/perl-5.10-uld/bin/perl on 64bit but curiously not with
111 /usr/local/perl-5.10-g/bin/perl. So it seems the behaviour is not even
112 in the test script always consistent.
114 * Todo: write a test that inserts a second dirty file with an already
115 existing timestamp. DONE
117 * Bug in perl 5.10 on my 64bit box:
119 DB<98> Devel::Peek::Dump $l
120 SV = PVMG(0x19e0450) at 0x142a550
122 FLAGS = (PADMY,NOK,POK,pNOK,pPOK)
124 NV = 1237123231.22458
125 PV = 0x194ce70 "1237123231.22458"\0
129 DB<99> Devel::Peek::Dump $r
130 SV = PVMG(0x19e0240) at 0x142a3e8
132 FLAGS = (PADMY,POK,pPOK)
135 PV = 0x19ff900 "1237123231.22458"\0
139 DB<100> Devel::Peek::Dump $l <=> $r
140 SV = IV(0x19ea6e8) at 0x19ea6f0
142 FLAGS = (PADTMP,IOK,pIOK)
145 DB<101> Devel::Peek::Dump $l
146 SV = PVMG(0x19e0450) at 0x142a550
148 FLAGS = (PADMY,NOK,POK,pNOK,pPOK)
150 NV = 1237123231.22458
151 PV = 0x194ce70 "1237123231.22458"\0
155 DB<102> Devel::Peek::Dump $r
156 SV = PVMG(0x19e0240) at 0x142a3e8
158 FLAGS = (PADMY,NOK,POK,pIOK,pNOK,pPOK)
160 NV = 1237123231.22458
161 PV = 0x19ff900 "1237123231.22458"\0
165 Retry with uselongdouble gives same effect. Not reproducable on 32bit box (k75).
167 * Todo: reset "done" or "covered" and "minmax" after a dirty operation?
169 2009-03-11 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
171 * $obj->merge ($other) needs to learn about equal epoch which may happen
172 since dirty_epoch intruded.
174 * Wontfix anytime soon: I think we currently do not support mkdir. Only
177 2009-01-01 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
179 * Todo: continue working on update(...,$dirty_epoch). It must be
180 followed by a fast_aggregate!
182 2008-12-26 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
184 * maybe we need a closest_entry or fitting_interval or something like
185 that. We want to merge an event into the middle of some recentfile.
186 First we do not know which file, then we do not know where to lock,
187 where to enter the new item, when and where to correct the dirtymark.
189 So my thought is we should first find which file.
191 Another part of my brain answers: what would happen if we would enter
192 the new file into the smallest file just like an ordinary new event,
193 just as an old event?
195 (1) we would write a duplicate timestamp? No, this would be easy to
198 (2) we would make the file large quickly? Yes, but so what? We are
199 changing the dirtymark, so are willing to disturb the downstream hosts.
201 2008-11-22 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
203 * 10705 root 17 0 725m 710m 1712 S 0.0 46.8 834:56.05 /home/src/perl/repoperls/installed-perls/perl/pVNtS9N/perl-5.8.0@32642/bin/perl -Ilib /home/k/sources/CPAN/GIT/trunk/bin/testing-rmirror.pl
207 https://rt.cpan.org/Ticket/Display.html?id=41199
209 * bzcat uploads.csv.bz2 | perl -F, -nale '$Seen{$F[-1]}++ and print'
211 Strangest output being HAKANARDO who managed to upload
213 Here is a better oneliner that includes also the first line of each
216 bzcat uploads.csv.bz2 | perl -MYAML::Syck -F, -nale '$F[-1]=~s/\s+\z//; push @{$Seen{$F[-1]}||=[]},$_; END {for my $k (keys %Seen){ delete $Seen{$k} if @{$Seen{$k}}==1; } print YAML::Syck::Dump(\%Seen)}'
218 2008-10-31 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
220 * memory leak in the syncher? It currently weighs 100M.
224 root 10705 1.0 4.9 80192 76596 pts/32 S+ Nov02 24:05 /home/src/perl/repoperls/installed-perls/perl/pVNtS9N/perl-5.8.0@32642/bin/perl -Ilib /home/k/sources/CPAN/GIT/trunk/bin/testing-rmirror.pl
227 2008-10-29 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
229 * lookup by epoch and by path and use this ability on the pause to never
230 again register a file twice that doesn't need it. Let's call it
233 * after the dirtymark is done: fill up recentfiles with fake (historic)
234 entries; fill up with individual corrections; algorithm maybe to be done
235 with bigfloat so that we can always place something in the middle
236 between two entries. Before we must switch to bigfloat we could try to
237 use Data::Float::nextup to get the.
239 * Inotify2 on an arbitrary tree and then play with that instead of PAUSE
242 * dirtymark now lives in Recentfile, needs to be used in rmirror.
244 * find out why the downloader died after a couple of hours without a net
245 connection. Write a test that survives the not-existence of the other
248 2008-10-15 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
250 * reconsider the HTTP epoch only. Not the whole thing over HTTP because
251 it makes less sense with tight coupling for secondary files. But asking
252 the server what the current epoch is might be cheaper on HTTP than on
253 rsync. (Needs to be evaluated)
255 * remove the 0.00 from the verbose overview in the Merged column in the
258 * write tests that expose the problems of the last few days: cascading
259 client/server roles, tight coupling for secondary RFs, deletes after
262 * Some day we might want to have policy options for the slave:
263 tight/loose/no coupling with upstream for secondary RFs. tight is what
264 we have now. loose would wait until a gap occurs that can be closed.
266 2008-10-14 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
268 * revisit all $rfs->[$i+1] places if they now make sense still
270 2008-10-11 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
272 * another bug is the fact that the mirror command deletes files before
273 it unhides the index file, thus confusing downstream slaves. We must not
274 delete before unhiding and must delete after unhiding. FIXED.
276 * new complication about the slave that is playing a server role.
277 Currently we mirror from newest to oldest with a hidden temporary file
278 as index. And when one file is finished, we unhide the index file.
279 Imagine the cascading server/slave is dead for a day. It then starts
280 mirroring again with the freshest thing and unhides the freshest index
281 file when it has worked through it. In that moment it exposes a time
282 hole. Because it now works on the second recentfile which is still
285 We currently do nothing special to converge after such a drop out. At
286 least not intentionally and robustly and thought through.
288 The algorithm we use to seed the next file needs quite a lot of more
289 robustness than it currently has. Something to do with looking at the
290 merged element of the next rf and when it has dropped off, we seed
291 immediately. And if it ramains dropped off, we seed again, of course.
293 Nope, looking from smaller to larger RFS we look at the merged element
294 of this RF and at the minmax/max element of the next RF. If that
295 $rf[next]->{minmax}{max} >= $rf[this]->{merged}{epoch}, then we can stop
298 And we need a public accessor seed and unseed or seeded. But not the mix
299 of public and private stuff that then is used behind the back.
301 And then the secondary* stuff must go.
303 And we must understand what the impact is on the DONE system. Can it go
304 unnoticed that there was a hole? And could the DONE system have decided
305 the hole is covered? This should be testable with three directories where
306 the middle stops working for a while. Done->merge is suspicious, we must
307 stop it from merging non-conflatable neighbors due to broken continuity.
311 2008-10-10 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
313 * Slaven suggests to have the current epoch or the whole current
314 recentfile available from the HTTP server and take it away with
315 keepalive. This direction goes the granularity down to subseconds.
317 We might want to rewrite everything to factor out transport and allow
318 the whole thing to run via HTTP.
320 2008-10-09 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
322 * are we sure we do NOT LEAVE DOT FILES around? Especially on the
325 * smoker on k81 fetching from k75 to verify cascading works. See
326 2008-07-17 in upgradexxx and rsync-over-recentfile-3.pl.
328 * maybe the loop should wait for CHECKSUMS file after every upload. And
329 CPAN.pm needs to deal with timestamps in the future.
331 * do not forget the dirtymark!
333 Text: have a new flag on recentfiles with the meaning: if this
334 changes, you're required to run a full rsync over all the files. The
335 reason why we set it would probably be: some foul happened. we injected
336 files in arbitrary places or didn't inject them although they changed.
337 The content of the flag? Timestamp? The relation between the
338 recentfiles would have to be inheritance from the principal, because any
339 out of band changes would soon later propagate to the next recentfile.
341 By upping the flag often one can easily ruin the slaves.
343 last out of band change? dirtymark?
345 Anyway, this implies that we read a potentially existing recentfile
348 And it implies that we have an eventloop that keeps us busy in 2-3
349 cycles, one for current stuff (tight loop) and one for the recentfiles
350 (cascade when principal has changed), one for the old stuff after a
353 And it implies that the out-of-band change in any of the recentfiles
354 must have a lock on the principal file and there is the place to set the
357 * start a FAQ, especially quick start guide questions. Also to aid those
358 problematic areas where we have no good solution, like the "links"
361 * wish feedback when we are slow.
365 * Remove a few DEBUG statements.
367 * The multiple-rrr way of doing things needs a new option to rmirror,
368 like piecemeal or so. Not urgent because after the first pass through,
369 things run smoothely. It's only ugly during the first pass.
371 * I have the suspicion that the code is broken that decides if the
372 neighboring RF needs to be seeded. I fear when too much time has gone
373 between two calls (in our case more than one hour), it would not seed
374 the neighbor. Of course this will never be noticed, so we need a good
377 * local/localroot confusion: I currently pass both options but one must
380 * accounts for early birds on PAUSE rsync daemon.
382 * hardcoded 20 seconds
384 * who mirrors the index? DOING now.
386 * which CPAN mirrors offer rsync?
388 * visit all XXX, visit all _float places
390 * rename the pathdb stuff, it's too confusing. No idea how.
392 * rrr-inotify, backpan, rrr-register
394 2008-10-08 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
396 * current bugs: the pathdb seems to get no reset, the seeding of the
397 secondaryttl stuff seems not to have an effect. Have helped myself with
398 a rand(10), need to fix this back. So not checked in. Does the rand
401 The rand thing helps. The secondaryttl stuff was in the wrong line,
404 The pathdb stuff was because I called either _pathdb or __pathdb on the
405 wrong object. FIXED now.
407 * It's not so beautiful if we never fetch the recentfiles that are not
408 the principal, even if this is correct behaviour. We really do not need
409 them after we have fetched the whole content.
411 OK, we want a switch for that: secondaryttl DONE
413 2008-10-07 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
415 * bug: rrr-news --max does not count correctly. with "35" it shows me 35
416 lines but with 36 it shows 110. First it repeats 35, gives 70, and then
417 it lets 40 follow. FIXED
419 * See that the long running process really only updates the principal
420 file unless it has missed a timespan during which something happened. If
421 nothing happened, it must notice even when it misses the timespan. DONE
423 * we must throw away the pathdb when we have reached the end of Z. From
424 that moment we can have a very small pathdb because the only reason for
425 a pathdb is that we know to ignore old records in old files. We won't
426 need this pathdb again before the next full pass over the data is
427 necessary and then we will rebuild it as we go along. DONE
429 2008-10-06 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
431 * I think, Done::register_one is doing wrong in that it does not
432 conflate neighboring pieces. The covered() method cannot do this because
433 it has no recent_events array at hand. But register_one has it and could
434 do it and for some reason misses to do it (sometimes).
436 This means that the three tests I just wrote can probably not survive
437 because they test with an already broken Done structure.
439 The art now is to detect how it happens, then to reproduce, then write a
442 So from the logfile this is what happens: we have a good interval with
443 newest file being F1 at T1. Now remotely F1 gets a change and F2 goes on
444 top of it. Locally we now mirror F2 and open a new done interval for it.
445 Then we mirror F1 but this time with the timestamp T1b. And when we then
446 try to close the gap, we do not find T1 but instead something older. We
447 should gladly accept this older piece and this would fix this bug.
451 * bug to fix: when the 1h file changes while rmirror is running, we do
452 correctly sync the new files but never switch to the 6h file but rather
453 stay in a rather quick loop that fetches the 1h file again and again.
455 Is it possible that we initialize a new object? Or does
456 get_remote_recentfile_as_tempfile overwrite something in myself?
458 Want a new option: _runstatusfile => $file which frequently dumps the
459 state of all recentfiles to a file.
463 2008-10-04 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
465 * Todo: now teach update to verify the timestamp is about to write
466 against the previous and use _increase_a_bit if it doesn't comply with
467 strict monotony. DONE
469 * The problem of rounding. So far perl's default precision was
470 sufficient. One day it won't be. FakeFloat has an easy job when it is
471 only reading and other machines have written correctly. But when we want
472 to write a floating point number that is a bit larger than the other
473 one, then we need our own idea of precision.
475 Slaven said: just append a "1". This might be going towards the end of
476 usability too quickly. I'd like something that actually uses the decimal
477 system. Well, appending a 1 also does this but...
479 E.g. we have 1.0. nextup on this architecture is starting with
480 1.0000000000000004. So there is a gap to fill: 1,2,3. Now I have
481 taken the 1.0000000000000003 and the next user comes and the time tells
482 him 1.0 again. He has to beat my number without stepping over the
483 nextup. This is much less space than I had when I chose 1,2,3.
485 What is also irritating is that nextup is architecture dependent. The
486 128 bit guy must choose very long numbers to fit in between whereas the
487 other one with 16 bit uses larger steps. But then the algorithm is the
488 same for both, so that would be a nice thing.
490 I see two situation where we need this. One is when Time::HiRes returns
491 us a value that is <= the last entry in our recentfile. In this case
492 (let's call it the end-case) we must fill the region between that number
493 and the next higher native floating point number. The other is when we
494 inject an old file into an old recentfile (we would then also set a new
495 dirtymark). We find the integer value already taken and need a slightly
496 different one (let's call it the middle-case). The difference between
497 the two situations is that the next user will want to find something
498 higher than my number in the end-case and something lower than my number
501 So I suggest we give the function both a value and an upper bound and it
502 calculates us a primitive middle. The upper bound in the middle-case is
503 the next integer. The upper bound on the end-case is the nextup floating
504 point number. But the latter poses another problem: if we have occupied
505 the middle m between x and nextup(x), then the nextup(m) will probably
506 not be the same as nextup(x) because some rounding will take place
507 before the nextup is calculated and when the rounding reaches the
508 nextup(x), we will end up at nextup(nextup(x)).
510 So we really need to consider the nextup and the nextdown from there and
511 then the middle and that's the number we may approach asymptotically.
514 2008-10-03 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
516 * consider deprecating the use of RECENT.recent as a symlink. It turns
517 out to need extra hoops with the rsync options and just isn't worth it.
518 Or maybe these extra hoops are needed anyway for the rest of the tree?
519 Nope, can't be the case because not all filesystems support symlinks.
521 But before doing the large step, I'll deprecate the call of
522 get_remote_recentfile_as_tempfile with an argument. Rememberr this was
523 only introduced to resolve RECENT.recent and complicates the routine far
524 beyond what it deserves.
526 DONE. Won't deprecate RECENT.recent, just moved its handling to the
529 2008-10-02 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
531 * I think it's a bug that the rsync_option links must be set to true in
532 order to support RECENT.recent and that nobody cares to set it
533 automatically. Similar for ignore_link_stat_errors.
535 2008-09-27 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
537 * Todo: find all todos together and make a plan what is missing for a
540 - verifytree or something like that. fsck maybe.
542 - rersyncrecent, the script itself? What it do?
544 - a way to only mirror the recentfiles without mirroring the whole
545 remote system such that people can decide to mirror only partially see
546 also 2008-08-30. .shadow-xxx directory? this also needed for a
547 filesystem that is still incomplete and might need the mirrorfiles for
550 - long living objects that mirror again and again. Inject something
551 into ta, see how it goes over to tb.
553 - how do we continue filling up the DONE system when we use an object
554 for the second time? "fully covered" and "uptodate" or new terminology.
556 - overview called on the wrong file should be understandable
558 - the meta data field that must change when we fake something up so that
559 the downstream people know they have to re-fetch everything.
561 - how tolerant are we against missing files upstream? how do we keep
562 track? there are legitimate cases where we did read upstream index right
563 before a file got deleted there and then find that file as new and want
564 it. There are other cases that are not self healing and must be tracked
567 - how, exactly, do we have to deal with deletes? With rsync errors?
569 rsync: link_stat "/id/K/KA/KARMAN/Rose-HTMLx-Form-Related-0.07.meta" (in
570 authors) failed: No such file or directory (2)
572 The file above is a delete in 1h and a new in file 1M and the
573 delete in the locally running rmirror did not get propagated to the 1M
574 object. Bug. And the consequence is a standstill.
576 It seems that a slave that works with a file below the principal needs
577 to merge things all the way up to get rid of later deletes. Or keep
578 track of all deletes and skip them later. So we need a trackdeletes.pm
579 similar to the done.pm?
581 see also 2008-08-20 about spurious deletes that really have no add
582 counterpart and yet they are not wrong.
584 - consider the effect when resyncing the recentfile takes longer than
585 the time per loop. Then we never rsync any file. We need to diagnose
586 that and force an increase of that loop time. But when we later are fast
587 enough again because the net has recovered, then we need to switch back
588 to original parameters. ERm, no, it's enough to keep syncing at least
589 one file before refetching an index file.
591 - remember to verify that no temp files are left lying around and the
594 - status file for not long running jobs that want to track upstream with
597 - revisit all XXX _float areas and study Sub::Exporter DONE
599 - persistent DB even though we just said we do not need it. Just for
600 extended capabilities and time savings when, for example, upstream
601 announces a reset and we get new recentfiles and could then limit
602 ourselves to a subset of files (those that have a changed epoch) in a
603 first pass and would only then do the loop to verify the rest. Or
606 * Todo: aggregate files should know their feed and finding the principal
607 should be done stepwise. (?)
609 * Todo: DESTROY thing that unlocks. Today when I left the debuggerr I
610 left locks around. DONE
612 2008-09-26 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
614 * maybe extend the _overview so that it always says if and where the
615 last file is in the next file and where the next event in the next rf
616 would lie. No, don't like this anymore. REJECT
618 * take the two new redundant tests out again, only the third must
621 * Todo: add a sanity check if the merged structure is really pointing to
622 a different rf and that this different rf is larger. DONE
624 2008-09-25 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
626 * now test, if they are overlapping. And test if there is a file in the
627 next rf that would fit into this rf's interval.
629 1h 1222324012.8474 1222322541.7963 0.4086
630 6h 1222320411.2760 1222304207.6931 4.5010 missing overlap/gap!
631 1d 1222320411.2760 1222238750.5071 22.6835 large overlap
632 1W 1222313218.3626 1221708477.5829 167.9835
634 I suspect that somebody writes a merged timestamp without having merged
635 and then somebody else relies on it.
637 If aggregate is running, the intervals must not be extravagated, if it
638 is not running, there must not be bounds, the total number of events in
639 the system must be counted and must be controlled throughout the tests.
640 That the test required the additional update was probably nonsense,
641 because aggregate can cut pieces too. FIXED & DONE
643 2008-09-23 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
645 * rrr-aggregate seems to rewrite the RECENT file even if nothing has
648 2008-09-21 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
650 * Most apparent bug at the moment is that the recentfiles are fetched
651 too often. Only the principal should be fetched and if it has not
652 changed, the others should not be refetched. ATM I must admit that I'm
653 happy that we refetch more often than needed because I can more easily
654 fix bugs while the thing is running.
656 * Let's say, 1220474966.19501 is a timestamp of a file that is already
657 done but the done system does not know about it. The reason for the
658 failure is not known and we never reach the status uptodate because of
659 this. We must get over it.
661 Later it turns out that the origin server had a bug somewhere.
662 1220474966.19042 came after 1220474966.19501. Or better: it was in the
663 array of the recentfile one position above. The bug was my own.
665 2008-09-20 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
667 * There is the race condition where the server does a delete and the
668 slave does not yet know and then tries to download it because he sees
669 the new. So for this time window we must be more tolerant against
670 failure. If we cannot download a file, we should just skip it and should
671 not retry immediately. The whole system should discover the lost thing
672 later. Keeping track with the DONE system should really be a no brainer.
674 But there is something more: the whole filesystem is a database and the
675 recentfiles are one possible representation of it. It's a pretty useful
676 representation I think that's why I have implemented something around
677 it. But for strictly local operation it has little value. For local
678 operation we would much rather have a database. So we would enter every
679 recentfile reading and every rsync operation and for every file the last
680 state change and what it leads to. Then we would always ignore older
681 records without the efforts involved with recentfiles.
683 The database would have: path,recentepoch,rsyncedon,deletedon
685 Oh well, not yet clear where this leads to.
687 2008-09-19 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
689 * Bug: the bigloop ran into a funny endless loop after EWILHELM uploaded
690 Module-Build. It *only* rsynced the "1h" recentfile from that moment on.
692 * statusfile, maybe only on demand, alone to have a sharp debugging
693 tool. It is locked and all recentfiles dump themselves into it and we
694 can build a viewer that lets us know where we stand and what's inside.
696 * remember: only the principal recentfile needs expiration, all others
697 shall be expired by principal if it discovers that something has move
700 2008-09-18 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
702 * Always check if we stringify to a higher value than in the entry
705 * And in covered make an additional check if we would be able to see a
706 numerical difference between the two numbers and if we can't then switch
707 to a different, more expensive algorithm. Do not want to be caught by
708 floating surprises. DONE
710 2008-09-17 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
712 * caching has several aspects here: we can cache the interval of the
713 recentfile which only will change when the mtime of the file changes. We
714 must re-mirror the recentfile when its ttl has expired. Does have_read
715 tell you anything? It counts nothing at all. Only the mtime is
716 interesting. The ntuple mtime, low-epoch, high-epoch. And as a separate
717 thing the have_mirrored because it is unrelated to the mtime.
719 * Robustness of floating point calculations! I always thought that the
720 string calculated by the origin server for the floating representation
721 of the epoch time is just a string. When we convert it to a number and
722 later back to a string, the other computer might come to a different
723 conclusion. This must not happen, we want to preserve it under any
724 circumstances. I will have to write tests with overlong sequences that
725 get lost in arithmetic and must see if all still works well. DONE
727 But one fragile point remains: if one host considers a>b and the other
728 one considers them == but no eq. To prevent this, we must probably do
729 some extra homework. DONE
731 2008-09-16 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
733 * the concept of tracking DONE needs an object per recentfile that has
734 something like these methods:
736 do_we_have(xxx), we_have(xxx), do_we_have_all(xxx,yyy), reset()
738 covered() register() covered()
740 The unclear thing is how we translate points in time into intervals. We
741 could pass a reference to the current recent_events array when running
742 we_have(xxx) and let the DONE object iterate over it such that it only
743 has to store a list of intervals that can melt into each other. Ah, even
744 passing the list together with a list of indexes seems feasiable.
746 Or maybe ask for the inverted list?
748 Whenever the complete array is covered by the interval we say we are
749 fully covered and if the recentfile is not expired, we are uptodate.
751 2008-09-07 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
753 2008-09-05 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
755 * need a way to "return" the next entry after the end of a list. When
756 the caller says "before" or "after" we would like to know if he could
757 cover that interval/threshold or not because this influences the effect
758 of a newer timestamp of that recentfile. DONE with $opt{info}.
760 2008-09-04 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
762 * one of the next things to tackle: the equivalent of csync2 -TIXU.
764 loop implies tixu (?). Nope, something like --statefile decides. Per
767 T test, I init, X including removals, U nodirtymark
769 So we have no concept of dirtymarks, we only trust that since we are
770 running we have observed everything steadily. But people will not let
771 this program run forever so we must consider both startup penalty and
772 book keeping for later runs. We keep this for later. For now we write a
773 long running mirror that merges several intervals.
775 2008-09-02 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
777 * need to speed up the 02 test, it's not clever to sleep so much. Reduce
780 * rersyncrecent, the script: default to one week. The name of the switch
781 is --after. Other switches? --loop!
783 2008-08-30 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
785 * need a switch --skip-deletes (?)
787 * need a switch --enduser that tells us that the whole tempfile
788 discipline is not needed when there is no downstream user. (?)
790 Without this switch we cannot have a reasonable recent.pl that just
791 displays the recent additions. Either we accept to download everything.
792 Or we download temporary files without the typical rsync protocol
795 Or maybe the switch is --tmpdir? If --tmpdir would mean: do not use
796 File::Temp::tempdir, this might be a win.
798 2008-08-29 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
800 * apropos missing: we have no push, we never know the downstream
801 servers. People who know their downstream hosts and want to ascertain
802 something will want additional methods we have never thought about, like
803 update or delete a certain file.
805 2008-08-26 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
807 * tempted to refactor rmirror into resolve_symlink, localize, etc.
808 Curious if rsync_options=links equal 0 vs. 1 will make the expected
811 * rsync options: it's a bit of a pain that we usually need several rsync
812 options, like compress, links, times, checksum and that there is no
813 reasonable default except the original rsync default. I think wee can
814 safely assume that the rsync options are shared between all recentfile
815 instances within one recent tree.
817 2008-08-20 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
819 * deletes: if a delete follows an add quickly enough it may happen that
820 a downstream mirror did not see the add at all! It seems this needs to
821 be mentioned somewhere. The point here is that even if the downstream is
822 never missing the principal timeframe it may encounter a "delete" that
823 has no complimentary "add" anywhere.
825 2008-08-19 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
827 * I suspect the treat of metadata is incorrect during read or something.
828 The bug that I am watching is that between 06:08 and 06:09 the 6h file
829 contained more than 6 hours worth of data. At 06:08 we merged into the
830 1d file. We need to take snapshots of the 6h file over the course of an
831 hour or maybe only between XX:08 and XX:09? Nope, the latter is not
834 Much worse: watching the 1h file: right at the moment (at 06:35) it
835 covers 1218867584-1219120397 which is 70 hours.
837 Something terribly broken. BTW, 1218867584 corresponds to Sat Aug 16
838 08:19:44 2008, that is when I checked out last time, so it seems to be
839 aggregating and never truncating?
841 No, correct is: it is never truncating; but wrong is: it is aggregating.
842 It does receive a lot of events from time to time from a larger file.
843 Somehow a large file gets merged into the small one and because the
844 "meta/merged" attribute is missing, nobody is paying attention. I
845 believe that I can fix this by making sure that metadata are honoured
846 during read. DONE and test adjusted.
848 2008-08-17 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
850 * grand renaming plan
852 remotebase => remoteroot to fit well with localroot DONE
853 local_path() => localroot seems to me should already work DONE
854 recentfile_basename => rfilename no need to stress it has no slash DONE
856 filenameroot??? Doesn't seem too bad to me today. Maybe something like
857 kern? It would anyway need a deprecation cycle because it is an
858 important constructor.
860 * I like the portability that Data::Serializer brings us but the price
861 is that some day we might find out that it is slowing us a bit. We'll
864 2008-08-16 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
866 * should we not enter the interval of the principal (or the interval of
867 the merging file?) in every aggregated/merged file?
869 * we should aim at a first release and give up on thinking about
870 sanitizing stuff and zloop. Let's just admit that a full traditional
871 rsync is the only available sanitizer ATM. Otherwise it's complicated
872 stuff: sanitizing on the origin server, sanitizing on the slaves,
873 sanitizing forgotten files, broken timestamps, etc. Let's delay it and
874 get the basics out before this becomes a major cause for mess.
876 2008-08-13 Andreas Koenig <k@andreas-koenigs-computer.local>
878 * On OSes not supporting symlinks we expect that RECENT.recent contains
879 the contents of the principal recentfile. Actually this is identical on
880 systems supporting symlinks. Simple, what follows from that is that we
881 need to keep the serializer in the metadata because we cannot read it
882 from the filename, doesn't it? Of course not. It's a chicken and egg
883 problem. This leaves us with the problem to actually parse the
884 serialized data to find out in which format it is. So who can do the 4
885 or 5 magics we wanted to support? File::LibMagic?
887 2008-08-09 Andreas Koenig <k@andreas-koenigs-computer.local>
889 * remotebase and recentfile_basename are ugly names. Now that we need a
890 word for the shortest/principal/driving recentfile too we should do
893 localroot is good. rfile is good. local_path() is bad, local_path($path)
894 is medium, filenameroot() is bad, remotebase is bad, recentfile is
897 Up to now remotebase was the string that described the remote root
898 directory in rsync notation, like pause.perl.org::authors. And
899 recentfile_basename was "RECENT-1h.yaml".
901 2008-08-08 Andreas Koenig <k@andreas-koenigs-computer.local>
903 * The test that was added in today's checkin is a good start for a test
904 of rmirror. We should have more methods in Recent.pm: verify,
905 addmissingfiles. We should verify the current tree, then rmirror it and
906 then verifytree the copy. We could then add some arbitrary file and let
907 it be discovered by addmissingfiles, then rmirror again and then
908 verifytree the copy again.
910 Then we could start stealing from csync2 sqlite database [no port to
911 OSX!] and fill a local DB. And methods to compare the database with the
912 recentfiles. Our strength is that in principle we could maintain state
913 with a single float. We have synced up to 1234567890.123456. If the Z
914 file does not add new files all we have to do is mirror the new ones and
917 This makes it clear that we should extend current protocol and declare
918 that we cheat when we add files too late, just to help the other end
919 keeping track. Ah yes, that's what was meant when zloop was mentioned
922 Maybe need to revisit File::Mirror to help me with this task.
924 2008-08-07 Andreas Koenig <k@andreas-koenigs-computer.local>
926 * There must be an allow-me-to-truncate flag in every recentfile.
927 Without it one could construct a sequence of updates winning the locking
928 battle against the aggregator. Only if an aggregator has managed to
929 merge data over to the next level, truncating can be allowed. DONE with
932 2008-08-06 Andreas Koenig <k@andreas-koenigs-computer.local>
934 * We should probably guarantee that no duplicates enter the aggregator
937 2008-08-02 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
939 * To get merge operation faster would need a good benchmark test. What
940 02 spits out isn't reliable enough and is dominated by many other
943 commit 10176bf6b79865d4fe9f46e3857a3b8669fa7961
944 Author: Andreas J. Koenig <k@k75.(none)>
945 Date: Sat Aug 2 07:58:04 2008 +0200
949 commit 3243120a0c120aaddcd9b1f4db6689ff12ed2523
950 Author: Andreas J. Koenig <k@k75.(none)>
951 Date: Sat Aug 2 11:40:29 2008 +0200
953 there was a lot of trying but the effect is hardly measurable with
956 * overhead of connecting seems high. When setting
957 max_files_per_connection to 1 we see that.
959 2008-08-01 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
961 * 1217622571.0889 - 1217597432.86734 = 25138.2215600014
963 25138.2215600014/3600 = 6.98283932222261
965 It jumps into the eye that this is ~ 7 hours, not ~6, so there seems to
966 be a bug in the aggregator. FIXED
968 2008-07-27 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
970 * e.g. id/Y/YE/YEWENBIN/Emacs-PDE-0.2.16.tar.gz: Do we have it, should
971 we have it, can we mirror it, mirror it!
973 I fear this needs a new class which might be called
974 File::Rsync::Mirror::Recent. It would collect all recentfiles of a kind
975 and treat them as an entity. I realize that a single recentfile may be
976 sufficient for certain tasks and that it is handy for the low level
977 programmer but it is not nice to use. If there is a delete in the 1h
978 file then the 6h file still contains it. Seekers of the best information
979 need to combine at least some of the recentfiles most of the time.
981 There is the place for the Z loop!
983 But the combination is something to collect in a database, isn't it. Did
984 csync2 just harrumph?
986 2008-07-26 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
988 * it just occurred to me that hosts in the same mirroring pool could
989 help out each other even without rewriting the recentfile. Just fetch
990 the stuff to mirror from several places, bingo. But that's something
991 that should rather live in a separate package or in rsync directly.
993 * cronjobs are unsuited because with ntp they would all come at the full
994 minute and disturb each other. Besides that I'd hate to have a backbone
995 with more than a few seconds latency.
997 2008-07-25 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
999 * a second rsync server with access control for PAUSE. Port? 873 is the
1000 standard port, let's take 8873.
1002 * if there were a filesystem based on this, it would have a slow access
1003 to inexistent files. It would probably provide wrong readdir (only based
1004 on current content) or also a slow one (based on a recentfile written
1005 after the call). But it would provide fast access to existing files. Or
1006 one would deliberately allow slightly blurred answers based on some
1007 sqlite reflection of the recentfiles.
1009 * todo: write a variant of mirror() that combines two or more
1010 recentfiles and treats them like one
1012 * todo: signal handler to remove the tempfile
1014 2008-07-24 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
1016 * now that we have the symlink I forgot how it should be used in
1019 * the z loop: add missing files to Z file. Just append them (instead of
1020 prepending). So one guy prepends something from the Y file from time to
1021 time and another guy appends something rather frequently. Collecting
1022 pond. When Y merges into Z, things get epoch and the collecting pond
1023 gets smaller. What exactly are "missing files"?
1025 take note of current epoch of the alpha file, let's call it the
1028 find all files on disk
1030 remove all files registered in the recentworld up to recent-ts
1032 remove all files that have been deleted after recent-ts according to
1035 2008-07-23 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
1037 * rersyncrecent might be a cronjob with a (locked) state file which
1038 contains things like after and maybe last z sync or such?
1040 rrr-mirror might be an alternative name but how would we justify the
1041 three Rs when there is no Re-Rsync-Recent?
1043 With the --loop parameter it is an endless loop, without it is no loop.
1044 At least this is simple.
1046 * todo: new accssor z-interval specifies how often the Z file is updated
1047 against the filesystem. We probably want no epoch stamp on these
1048 entries. And we want to be able to filter the entries (e.g. no
1049 by-modules and by-category tree)
1051 2008-07-20 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
1053 * Fill the Z file. gc or fsck or both. Somehow we must get the old files
1054 into Z. We do not need the other files filled up with filesystem
1057 * need interface to query for a file in order to NOT call update on
1058 PAUSE a second time within a short time.
1060 2008-07-19 Andreas J. Koenig <andreas.koenig.7os6VVqR@franz.ak.mind.de>
1062 * recommended update interval? Makes no sense, is different for
1069 change-log-default-name: "Todo"