round() is from C99. Use rint() instead. There are behavioral
differences between round() and rint(), but they should not matter to
the Bloom filter optimal_k() function. We already assume POSIX
behavior for rint(), so there is no question of rint() not using
"rounds towards nearest" as its rounding mode.
Cleanup from commit 51bc271790.
Per buildfarm member thrips.
Author: Peter Geoghegan
Discussion: https://postgr.es/m/CAH2-Wzn76eCGUonARy-wrVtMHsf+4cvbK_oJAWTLfORTU5ki0w@mail.gmail.com
A Bloom filter is a space-efficient, probabilistic data structure that
can be used to test set membership. Callers will sometimes incur false
positives, but never false negatives. The rate of false positives is a
function of the total number of elements and the amount of memory
available for the Bloom filter.
Two classic applications of Bloom filters are cache filtering, and data
synchronization testing. Any user of Bloom filters must accept the
possibility of false positives as a cost worth paying for the benefit in
space efficiency.
This commit adds a test harness extension module, test_bloomfilter. It
can be used to get a sense of how the Bloom filter implementation
performs under varying conditions.
This is infrastructure for the upcoming "heapallindexed" amcheck patch,
which verifies the consistency of a heap relation against one of its
indexes.
Author: Peter Geoghegan
Reviewed-By: Andrey Borodin, Michael Paquier, Thomas Munro, Andres Freund
Discussion: https://postgr.es/m/CAH2-Wzm5VmG7cu1N-H=nnS57wZThoSDQU+F5dewx3o84M+jY=g@mail.gmail.com
In 9956ddc191, ten years ago, the
current objfile.txt based linking model was introduced. It's time to
retire the old SUBSYS.o based model.
This primarily is pertinent because the bitcode files for LLVM based
inlining are not produced when using PARTIAL_LINKING. It does not seem
worth to fix PARTIAL_LINKING to support that.
Author: Andres Freund
Discussion: https://postgr.es/m/20180121204356.d5oeu34jetqhmdv2@alap3.anarazel.de
LockViewRecurese() obtains view relation using heap_open() and passes
it to get_view_query() to get view info. It immediately closes the
relation then uses the returned view info by calling
LockViewRecurse_walker(). Since get_view_query() returns a pointer
within the relcache, the relcache should be kept until
LockViewRecurse_walker() returns. Otherwise the relation could point
to a garbage memory area.
Fix is moving the heap_close() call after LockViewRecurse_walker().
Problem reported by Tom Lane (buildfarm is unhappy, especially prion
since it enables -DRELCACHE_FORCE_RELEASE cpp flag), fix by me.
A followup patch will add a SKIP_LOCKED option. To avoid introducing
evermore arguments, breaking existing callers each time, introduce a
flags argument. This'll no doubt break a few external users...
Also change the MISSING_OK behaviour so a DEBUG1 debug message is
emitted when a relation is not found.
Author: Nathan Bossart
Reviewed-By: Michael Paquier and Andres Freund
Discussion: https://postgr.es/m/20180306005349.b65whmvj7z6hbe2y@alap3.anarazel.de
Previously there was no way in the standby side to find out the host and port
of the sender server that the walreceiver was currently connected to when
multiple hosts and ports were specified in primary_conninfo. For that purpose,
this patch adds sender_host and sender_port columns into pg_stat_wal_receiver
view. They report the host and port that the active replication connection
currently uses.
Bump catalog version.
Author: Haribabu Kommi
Reviewed-by: Michael Paquier and me
Discussion: https://postgr.es/m/CAJrrPGcV_aq8=cdqkFhVDJKEnDQ70yRTTdY9RODzMnXNrCz2Ow@mail.gmail.com
In the previous coding, skipped pages were mostly zeroes, but they still
had valid WAL page headers. That makes them very much less compressible
than an unbroken string of zeroes would be --- about 10X worse for bzip2
compression, for instance. We don't need those headers, so tweak the logic
so that we zero them out.
Chapman Flack, reviewed by Daniel Gustafsson
Discussion: https://postgr.es/m/579297F8.7020107@anastigmatix.net
When SSI was developed, slru.c was limited to segment files with names in
the range 0000-FFFF. This didn't allow enough space for predicate.c to
store every possible XID when spilling old transactions to disk, so it
would wrap around sooner and print warnings. Since commits 638cf09e and
73c986ad increased the number of segment files slru.c could manage, that
behavior is unnecessary. Therefore remove that code.
Also remove the macro OldSerXidSegment, which has been unused since
4cd3fb6e.
Thomas Munro, reviewed by Anastasia Lubennikova
Discussion: https://postgr.es/m/CAEepm=3XfsTSxgEbEOmxu0QDiXy0o18NUg2nC89JZcCGE+XFPA@mail.gmail.com
Add the target context's name to the errdetail field of "out of memory"
errors in mcxt.c. Per discussion, this seems likely to be useful to
help narrow down the cause of a reported failure, and it costs little.
Also, now that context names are required to be compile-time constants
in all cases, there's little reason to be concerned about security
issues from exposing these names to users. (Because of such concerns,
we are *not* including the context "ident" field.)
In passing, add unlikely() markers to the allocation-failed tests,
just to be sure the compiler is on the right page about that.
Also, in palloc and friends, copy CurrentMemoryContext into a local
variable, as that's almost surely cheaper to reference than a global.
Discussion: https://postgr.es/m/1099.1522285628@sss.pgh.pa.us
In btree and SP-GiST indexes, move the responsibility for calling
IndexFreeSpaceMapVacuum from the vacuumcleanup phase to the bulkdelete
phase, and do it if and only if we found some pages that could be put into
FSM. As in commit 851a26e26, the idea is to make free pages visible to FSM
searchers sooner when vacuuming very large tables (large enough to need
multiple bulkdelete scans). This adds more redundant work than that commit
did, since we have to scan the entire index FSM each time rather than being
able to localize what needs to be updated; but it still seems worthwhile.
However, we can buy something back by not touching the FSM at all when
there are no pages that can be put in it. That will result in slower
recovery from corrupt upper FSM pages in such a scenario, but it doesn't
seem like that's a case we need to optimize for.
Hash indexes don't use FSM at all. GIN, GiST, and bloom indexes update
FSM during the vacuumcleanup phase not bulkdelete, so that doing something
comparable to this would be a much more invasive change, and it's not clear
it's worth it. BRIN indexes do things sufficiently differently that this
change doesn't apply to them, either.
Claudio Freire, reviewed by Masahiko Sawada and Jing Wang, some additional
tweaks by me
Discussion: https://postgr.es/m/CAGTBQpYR0uJCNTt3M5GOzBRHo+-GccNO1nCaQ8yEJmZKSW5q1A@mail.gmail.com
Unlike the previous coding, this might result in a Gather per Append
subplan when the target list is parallel-restricted, but such a plan
is probably worth considering in that case, since a single Gather
on top of the entire Append is impossible.
Per Andres Freund and the buildfarm.
Discussion: http://postgr.es/m/20180330050351.bmxx4cdtz67czjda@alap3.anarazel.de
Predicate locks are used on per page basis only if fastupdate = off, in
opposite case predicate lock on pending list will effectively lock whole index,
to reduce locking overhead, just lock a relation. Entry and posting trees are
essentially B-tree, so locks are acquired on leaf pages only.
Author: Shubham Barai with some editorization by me and Dmitry Ivanov
Review by: Alexander Korotkov, Dmitry Ivanov, Fedor Sigaev
Discussion: https://www.postgresql.org/message-id/flat/CALxAEPt5sWW+EwTaKUGFL5_XFcZ0MuGBcyJ70oqbWqr42YKR8Q@mail.gmail.com
If the toplevel scan/join target list is parallel-safe, postpone
generating Gather (or Gather Merge) paths until after the toplevel has
been adjusted to return it. This (correctly) makes queries with
expensive functions in the target list more likely to choose a
parallel plan, since the cost of the plan now reflects the fact that
the evaluation will happen in the workers rather than the leader.
The original complaint about this problem was from Jeff Janes.
If the toplevel scan/join relation is partitioned, recursively apply
the changes to all partitions. This sometimes allows us to get rid of
Result nodes, because Append is not projection-capable but its
children may be. It also cleans up what appears to be incorrect SRF
handling from commit e2f1eb0ee30d144628ab523432320f174a2c8966: the old
code had no knowledge of SRFs for child scan/join rels.
Because we now use create_projection_path() in some cases where we
formerly used apply_projection_to_path(), this changes the ordering
of columns in some queries generated by postgres_fdw. Update
regression outputs accordingly.
Patch by me, reviewed by Amit Kapila and by Ashutosh Bapat. Other
fixes for this problem (substantially different from this version)
were reviewed by Dilip Kumar, Amit Khandekar, and Marina Polyakova.
Discussion: http://postgr.es/m/CAMkU=1ycXNipvhWuweUVpKuyu6SpNjF=yHWu4c4US5JgVGxtZQ@mail.gmail.com
Don't call generate_gather_paths for the topmost scan/join relation
when it is initially populated with paths. Instead, do the work in
grouping_planner. By itself, this gains nothing; in fact it loses
slightly because we end up calling set_cheapest() for the topmost
scan/join rel twice rather than once. However, it paves the way for
a future commit which will postpone generate_gather_paths for the
topmost scan/join relation even further, allowing more accurate
costing of parallel paths.
Amit Kapila and Robert Haas. Earlier versions of this patch (which
different substantially) were reviewed by Dilip Kumar, Amit
Khandekar, Marina Polyakova, and Ashutosh Bapat.
We sometimes insert a ProjectionPath into a plan tree when projection
is not strictly required. The existing code already arranges to avoid
emitting a Result node when the ProjectionPath's subpath can perform
the projection itself, but previously it didn't consider the
possibility that the parent node might not actually require the
projection to be performed at all.
Skipping projection when it's not required can not only avoid Result
nodes that aren't needed, but also avoid losing the "physical tlist"
optimization unneccessarily.
Patch by me, reviewed by Amit Kapila.
Discussion: http://postgr.es/m/CA+TgmoakT5gmahbPWGqrR2nAdFOMAOnOXYoWHRdVfGWs34t6_A@mail.gmail.com
Just noticed that these were quite redundant, since we're holding the
page address in a local variable anyway, and we have pin on the buffer
throughout.
Also improve a comment.
FreeSpaceMapVacuumRange has the same effect, is more efficient if many
pages are involved, and makes fewer assumptions about how it's used.
Notably, Claudio Freire pointed out that UpdateFreeSpaceMap could fail
if the specified freespace value isn't the maximum possible. This isn't
a problem for the single existing user, but the function represents an
attractive nuisance IMO, because it's named as though it were a
general-purpose update function and its limitations are undocumented.
In any case we don't need multiple ways to get the same result.
In passing, do some code review and cleanup in RelationAddExtraBlocks.
In particular, I see no excuse for it to omit the PageIsNew safety check
that's done in the mainline extension path in RelationGetBufferForTuple.
Discussion: https://postgr.es/m/CAGTBQpYR0uJCNTt3M5GOzBRHo+-GccNO1nCaQ8yEJmZKSW5q1A@mail.gmail.com
VACUUM updates leaf-level FSM entries immediately after cleaning the
corresponding heap blocks. fsmpage.c updates the intra-page search trees
on the leaf-level FSM pages when this happens, but it does not touch the
upper-level FSM pages, so that the released space might not actually be
findable by searchers. Previously, updating the upper-level pages happened
only at the conclusion of the VACUUM run, in a single FreeSpaceMapVacuum()
call. This is bad because the VACUUM might get canceled before ever
reaching that point, so that from the point of view of searchers no space
has been freed at all, leading to table bloat.
We can improve matters by updating the upper pages immediately after each
cycle of index-cleaning and heap-cleaning, processing just the FSM pages
corresponding to the range of heap blocks we have now fully cleaned.
This adds a small amount of extra work, since the FSM pages leading down
to each range boundary will be touched twice, but it's pretty negligible
compared to everything else going on in a large VACUUM.
If there are no indexes, VACUUM doesn't work in cycles but just cleans
each heap page on first visit. In that case we just arbitrarily update
upper FSM pages after each 8GB of heap. That maintains the goal of not
letting all this work slide until the very end, and it doesn't seem worth
expending extra complexity on a case that so seldom occurs in practice.
In either case, the FSM is fully up to date before any attempt is made
to truncate the relation, so that the most likely scenario for VACUUM
cancellation no longer results in out-of-date upper FSM pages. When
we do successfully truncate, adjusting the FSM to reflect that is now
fully handled within FreeSpaceMapTruncateRel.
Claudio Freire, reviewed by Masahiko Sawada and Jing Wang, some additional
tweaks by me
Discussion: https://postgr.es/m/CAGTBQpYR0uJCNTt3M5GOzBRHo+-GccNO1nCaQ8yEJmZKSW5q1A@mail.gmail.com
Add explicit cast from scalar jsonb to all numeric and bool types. It would be
better to have cast from scalar jsonb to text too but there is already a cast
from jsonb to text as just text representation of json. There is no way to have
two different casts for the same type's pair.
Bump catalog version
Author: Anastasia Lubennikova with editorization by Nikita Glukhov and me
Review by: Aleksander Alekseev, Nikita Glukhov, Darafei Praliaskouski
Discussion: https://www.postgresql.org/message-id/flat/0154d35a-24ae-f063-5273-9ffcdf1c7f2e@postgrespro.ru
Previously, committing or aborting inside a cursor loop was prohibited
because that would close and remove the cursor. To allow that,
automatically convert such cursors to holdable cursors so they survive
commits or rollbacks. Portals now have a new state "auto-held", which
means they have been converted automatically from pinned. An auto-held
portal is kept on transaction commit or rollback, but is still removed
when returning to the main loop on error.
This supports all languages that have cursor loop constructs: PL/pgSQL,
PL/Python, PL/Perl.
Reviewed-by: Ildus Kurbangaliev <i.kurbangaliev@postgrespro.ru>
As promised in earlier commits, this adds documentation about the new
build options, the new GUCs, about the planner logic when JIT is used,
and the benefits of JIT in general.
Also adds a more implementation oriented README.
I'm sure we're going to want to expand this further, but I think this
is a reasonable start.
Author: Andres Freund, with contributions by Thomas Munro
Reviewed-By: Thomas Munro
Discussion: https://postgr.es/m/20170901064131.tazjxwus3k2w3ybh@alap3.anarazel.de
This just shows a few details about JITing, e.g. how many functions
have been JITed, and how long that took. To avoid noise in regression
tests with functions sometimes being JITed in --with-llvm builds,
disable display when COSTS OFF is specified.
Author: Andres Freund
Discussion: https://postgr.es/m/20170901064131.tazjxwus3k2w3ybh@alap3.anarazel.de
This provides infrastructure to allow JITed code to inline code
implemented in C. This e.g. can be postgres internal functions or
extension code.
This already speeds up long running queries, by allowing the LLVM
optimizer to optimize across function boundaries. The optimization
potential currently doesn't reach its full potential because LLVM
cannot optimize the FunctionCallInfoData argument fully away, because
it's allocated on the heap rather than the stack. Fixing that is
beyond what's realistic for v11.
To be able to do that, use CLANG to convert C code to LLVM bitcode,
and have LLVM build a summary for it. That bitcode can then be used to
to inline functions at runtime. For that the bitcode needs to be
installed. Postgres bitcode goes into $pkglibdir/bitcode/postgres,
extensions go into equivalent directories. PGXS has been modified so
that happens automatically if postgres has been compiled with LLVM
support.
Currently this isn't the fastest inline implementation, modules are
reloaded from disk during inlining. That's to work around an apparent
LLVM bug, triggering an apparently spurious error in LLVM assertion
enabled builds. Once that is resolved we can remove the superfluous
read from disk.
Docs will follow in a later commit containing docs for the whole JIT
feature.
Author: Andres Freund
Discussion: https://postgr.es/m/20170901064131.tazjxwus3k2w3ybh@alap3.anarazel.de
The target cluster that was rewound needs to perform recovery from
the checkpoint created at failover, which leads it to remove or recreate
some files and directories that may have been copied from the source
cluster. So pg_rewind can skip synchronizing such files and directories,
and which reduces the amount of data transferred during a rewind
without changing the usefulness of the operation.
Author: Michael Paquier
Reviewed-by: Anastasia Lubennikova, Stephen Frost and me
Discussion: https://postgr.es/m/20180205071022.GA17337@paquier.xyz
So far, a nested CALL or DO in PL/pgSQL would not establish a context
where transaction control statements were allowed. This fixes that by
handling CALL and DO specially in PL/pgSQL, passing the atomic/nonatomic
execution context through and doing the required management around
transaction boundaries.
Reviewed-by: Tomas Vondra <tomas.vondra@2ndquadrant.com>
tuplesort_gettupleslot() passed back tuples allocated in the tuplesort's
own memory context, even when the caller was responsible to free them.
This created a double-free hazard, because some callers might destroy
the tuplesort object (via tuplesort_end) before trying to clean up the
last returned tuple. To avoid this, change the API to specify that the
tuple is allocated in the caller's memory context. v10 and HEAD already
did things that way, but in 9.5 and 9.6 this is a live bug that can
demonstrably cause crashes with some grouping-set usages.
In 9.5 and 9.6, this requires doing an extra tuple copy in some cases,
which is unfortunate. But the amount of refactoring needed to avoid it
seems excessive for a back-patched change, especially since the cases
where an extra copy happens are less performance-critical.
Likewise change tuplesort_getdatum() to return pass-by-reference Datums
in the caller's context not the tuplesort's context. There seem to be
no live bugs among its callers, but clearly the same sort of situation
could happen in future.
For other tuplesort fetch routines, continue to allocate the memory in
the tuplesort's context. This is a little inconsistent with what we now
do for tuplesort_gettupleslot() and tuplesort_getdatum(), but that's
preferable to adding new copy overhead in the back branches where it's
clearly unnecessary. These other fetch routines provide the weakest
possible guarantees about tuple memory lifespan from v10 on, anyway,
so this actually seems more consistent overall.
Adjust relevant comments to reflect these API redefinitions.
Arguably, we should change the pre-9.5 branches as well, but since
there are no known failure cases there, it seems not worth the risk.
Peter Geoghegan, per report from Bernd Helmle. Reviewed by Kyotaro
Horiguchi; thanks also to Andreas Seltenreich for extracting a
self-contained test case.
Discussion: https://postgr.es/m/1512661638.9720.34.camel@oopsware.de
Store GID of 2PC in commit/abort WAL records when wal_level = logical.
This allows logical decoding to send the SAME gid to subscribers
across restarts of logical replication.
Track relica origin replay progress for 2PC.
(Edited from patch 0003 in the logical decoding 2PC series.)
Authors: Nikhil Sontakke, Stas Kelvich
Reviewed-by: Simon Riggs, Andres Freund
Instead using memset to set tts_isnull, call the new
slot_getmissingattrs().
Also fix a bug (= instead of >=) in the code generation. Normally = is
correct, but when repeatedly deforming fields not in a
tuple (e.g. deform up to natts + 1 and then natts + 2) >= is needed.
Discussion: https://postgr.es/m/20180328010053.i2qvsuuusst4lgmc@alap3.anarazel.de
Currently adding a column to a table with a non-NULL default results in
a rewrite of the table. For large tables this can be both expensive and
disruptive. This patch removes the need for the rewrite as long as the
default value is not volatile. The default expression is evaluated at
the time of the ALTER TABLE and the result stored in a new column
(attmissingval) in pg_attribute, and a new column (atthasmissing) is set
to true. Any existing row when fetched will be supplied with the
attmissingval. New rows will have the supplied value or the default and
so will never need the attmissingval.
Any time the table is rewritten all the atthasmissing and attmissingval
settings for the attributes are cleared, as they are no longer needed.
The most visible code change from this is in heap_attisnull, which
acquires a third TupleDesc argument, allowing it to detect a missing
value if there is one. In many cases where it is known that there will
not be any (e.g. catalog relations) NULL can be passed for this
argument.
Andrew Dunstan, heavily modified from an original patch from Serge
Rielau.
Reviewed by Tom Lane, Andres Freund, Tomas Vondra and David Rowley.
Discussion: https://postgr.es/m/31e2e921-7002-4c27-59f5-51f08404c858@2ndQuadrant.com
Originally, we treated memory context names as potentially variable in
all cases, and therefore always copied them into the context header.
Commit 9fa6f00b1 rethought this a little bit and invented a distinction
between fixed and variable names, skipping the copy step for the former.
But we can make things both simpler and more useful by instead allowing
there to be two parts to a context's identification, a fixed "name" and
an optional, variable "ident". The name supplied in the context create
call is now required to be a compile-time-constant string in all cases,
as it is never copied but just pointed to. The "ident" string, if
wanted, is supplied later. This is needed because typically we want
the ident to be stored inside the context so that it's cleaned up
automatically on context deletion; that means it has to be copied into
the context before we can set the pointer.
The cost of this approach is basically just an additional pointer field
in struct MemoryContextData, which isn't much overhead, and is bought
back entirely in the AllocSet case by not needing a headerSize field
anymore, since we no longer have to cope with variable header length.
In addition, we can simplify the internal interfaces for memory context
creation still further, saving a few cycles there. And it's no longer
true that a custom identifier disqualifies a context from participating
in aset.c's freelist scheme, so possibly there's some win on that end.
All the places that were using non-compile-time-constant context names
are adjusted to put the variable info into the "ident" instead. This
allows more effective identification of those contexts in many cases;
for example, subsidary contexts of relcache entries are now identified
by both type (e.g. "index info") and relname, where before you got only
one or the other. Contexts associated with PL function cache entries
are now identified more fully and uniformly, too.
I also arranged for plancache contexts to use the query source string
as their identifier. This is basically free for CachedPlanSources, as
they contained a copy of that string already. We pay an extra pstrdup
to do it for CachedPlans. That could perhaps be avoided, but it would
make things more fragile (since the CachedPlanSource is sometimes
destroyed first). I suspect future improvements in error reporting will
require CachedPlans to have a copy of that string anyway, so it's not
clear that it's worth moving mountains to avoid it now.
This also changes the APIs for context statistics routines so that the
context-specific routines no longer assume that output goes straight
to stderr, nor do they know all details of the output format. This
is useful immediately to reduce code duplication, and it also allows
for external code to do something with stats output that's different
from printing to stderr.
The reason for pushing this now rather than waiting for v12 is that
it rethinks some of the API changes made by commit 9fa6f00b1. Seems
better for extension authors to endure just one round of API changes
not two.
Discussion: https://postgr.es/m/CAB=Je-FdtmFZ9y9REHD7VsSrnCkiBhsA4mdsLKSPauwXtQBeNA@mail.gmail.com
If the value of an index expression is unchanged after UPDATE,
allow HOT updates where previously we disallowed them, giving
a significant performance boost in those cases.
Particularly useful for indexes such as JSON->>field where the
JSON value changes but the indexed value does not.
Submitted as "surjective indexes" patch, now enabled by use
of new "recheck_on_update" parameter.
Author: Konstantin Knizhnik
Reviewer: Simon Riggs, with much wordsmithing and some cleanup
Add page-level predicate locking, due to gist's code organization, patch seems
close to trivial: add check before page changing, add predicate lock before page
scanning. Although choosing right place to check is not simple: it should not
be called during index build, it should support insertion of new downlink and so
on.
Author: Shubham Barai with editorization by me and Alexander Korotkov
Reviewed by: Alexander Korotkov, Andrey Borodin, me
Discussion: https://www.postgresql.org/message-id/flat/CALxAEPtdcANpw5ePU3LvnTP8HCENFw6wygupQAyNBgD-sG3h0g@mail.gmail.com
This is mostly done to be able to validate features and fixes
submitted to LLVM. Given the size of these changes that seems
acceptable.
Author: Andres Freund
Due to the differing APIs between versions, I forgot to deallocate the
generated module in older LLVM versions, leading to a memory leak.
Author: Andres Freund
Performing JIT compilation for deforming gains performance benefits
over unJITed deforming from compile-time knowledge of the tuple
descriptor. Fixed column widths, NOT NULLness, etc can be taken
advantage of.
Right now the JITed deforming is only used when deforming tuples as
part of expression evaluation (and obviously only if the descriptor is
known). It's likely to be beneficial in other cases, too.
By default tuple deforming is JITed whenever an expression is JIT
compiled. There's a separate boolean GUC controlling it, but that's
expected to be primarily useful for development and benchmarking.
Docs will follow in a later commit containing docs for the whole JIT
feature.
Author: Andres Freund
Discussion: https://postgr.es/m/20170901064131.tazjxwus3k2w3ybh@alap3.anarazel.de
The listed numbers disagreed with the ones being used in the symbols;
but instead of just fixing the numbers in the comment, use the symbolic
name instead, which seems clearer.
This has been wrong all along, so apply back to 9.5 where BRIN was
introduced.
Reported-by: Tomas Vondra
Discussion: https://postgr.es/m/5ff514f2-8b1e-6366-b11c-8e2ed442562d@2ndquadrant.com
Andres Freund