Обсуждение: TRACE_SORT defined by default
I just noticed that TRACE_SORT is defined by default (since 2005 apparently). It seems odd since it is the only debugging code enabled by default. Joe -- Crunchy Data - http://crunchydata.com PostgreSQL Support for Secure Enterprises Consulting, Training, & Open Source Development
Вложения
On Wed, Apr 24, 2019 at 2:07 PM Joe Conway <mail@joeconway.com> wrote: > I just noticed that TRACE_SORT is defined by default (since 2005 > apparently). It seems odd since it is the only debugging code enabled by > default. I think that we should get rid of the #ifdef stuff, so that it isn't possible to disable the trace_sort instrumentation my commenting out the TRACE_SORT entry in pg_config_manual.h. I recall being opposed on this point by Robert Haas. Possibly because he just didn't want to deal with it at the time. -- Peter Geoghegan
On 4/24/19 5:10 PM, Peter Geoghegan wrote: > On Wed, Apr 24, 2019 at 2:07 PM Joe Conway <mail@joeconway.com> wrote: >> I just noticed that TRACE_SORT is defined by default (since 2005 >> apparently). It seems odd since it is the only debugging code enabled by >> default. > > I think that we should get rid of the #ifdef stuff, so that it isn't > possible to disable the trace_sort instrumentation my commenting out > the TRACE_SORT entry in pg_config_manual.h. I recall being opposed on > this point by Robert Haas. Possibly because he just didn't want to > deal with it at the time. Has anyone ever (or recently) measured the impact on performance to have this enabled? Is it that generically useful for debugging of production instances of Postgres that we really want it always enabled despite the performance impact? Maybe the answer to both is "yes", but if so I would agree that we ought to remove the define and ifdef's and just bake it in. Joe -- Crunchy Data - http://crunchydata.com PostgreSQL Support for Secure Enterprises Consulting, Training, & Open Source Development
Вложения
On Wed, Apr 24, 2019 at 2:15 PM Joe Conway <mail@joeconway.com> wrote: > Has anyone ever (or recently) measured the impact on performance to have > this enabled? Is it that generically useful for debugging of production > instances of Postgres that we really want it always enabled despite the > performance impact? It is disabled by default, in the sense that the trace_sort GUC defaults to off. I believe that the overhead of building in the instrumentation without enabling it is indistinguishable from zero. In any case the current status quo is that it's built by default. I have used it in production, though not very often. It's easy to turn it on and off. > Maybe the answer to both is "yes", but if so I would agree that we ought > to remove the define and ifdef's and just bake it in. We're only talking about removing the option of including the instrumentation in binaries when Postgres is built. I'm not aware that anyone is doing that. It nobody was doing that, then nobody could be affected by removing the #ifdef crud. I suspect that the reason that this hasn't happened already is because it leaves trace_sort/TRACE_SORT in the slightly awkward position of no longer quite meeting the traditional definition of a "developer option". -- Peter Geoghegan
On 2019-Apr-24, Peter Geoghegan wrote: > I suspect that the reason that this hasn't happened already is because > it leaves trace_sort/TRACE_SORT in the slightly awkward position of no > longer quite meeting the traditional definition of a "developer > option". This is a really strange argument. You're saying that somebody thought about it: "Hmm, well, I can remove this preprocessor symbol but then trace_sort would no longer resemble a developer option. So I'm going to leave the symbol alone". I don't think that's what happened. It seems more likely to me that nobody has gone to the trouble of deciding that the symbol is worth removing, let alone actually doing it. If the instrumentation is good, and you seem to be saying that it is, I think we should just remove the symbol and be done with it. -- Álvaro Herrera https://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
On Wed, Apr 24, 2019 at 2:29 PM Alvaro Herrera <alvherre@2ndquadrant.com> wrote: > This is a really strange argument. You're saying that somebody thought > about it: "Hmm, well, I can remove this preprocessor symbol but then > trace_sort would no longer resemble a developer option. So I'm going to > leave the symbol alone". I don't think that's what happened. It seems > more likely to me that nobody has gone to the trouble of deciding that > the symbol is worth removing, let alone actually doing it. It doesn't seem very important now. > If the instrumentation is good, and you seem to be saying that it is, I > think we should just remove the symbol and be done with it. Sounds like a plan. Do you want to take care of it, Joe? -- Peter Geoghegan
Peter Geoghegan <pg@bowt.ie> writes:
> On Wed, Apr 24, 2019 at 2:15 PM Joe Conway <mail@joeconway.com> wrote:
>> Has anyone ever (or recently) measured the impact on performance to have
>> this enabled? Is it that generically useful for debugging of production
>> instances of Postgres that we really want it always enabled despite the
>> performance impact?
> It is disabled by default, in the sense that the trace_sort GUC
> defaults to off. I believe that the overhead of building in the
> instrumentation without enabling it is indistinguishable from zero.
It would probably be useful to actually prove that rather than just
assuming it. I do see some code under the symbol that is executed
even when !trace_sort, and in any case Andres keeps complaining that
even always-taken branches are expensive ...
> In
> any case the current status quo is that it's built by default. I have
> used it in production, though not very often. It's easy to turn it on
> and off.
Would any non-wizard really have a use for it?
It seems like we should either make this really a developer option
(and hence not enabled by default) or else move it into some other
category than DEVELOPER_OPTIONS.
regards, tom lane
On Wed, Apr 24, 2019 at 3:04 PM Tom Lane <tgl@sss.pgh.pa.us> wrote: > > It is disabled by default, in the sense that the trace_sort GUC > > defaults to off. I believe that the overhead of building in the > > instrumentation without enabling it is indistinguishable from zero. > > It would probably be useful to actually prove that rather than just > assuming it. The number of individual trace_sort LOG messages is proportionate to the number of runs produced. > I do see some code under the symbol that is executed > even when !trace_sort, and in any case Andres keeps complaining that > even always-taken branches are expensive ... I think that you're referring to the stuff needed for the D-Trace probes. It's a pity that there isn't better support for that, since Linux has a lot for options around static userspace probes these days (SystemTap is very much out of favor, and never was very popular). There seems to be a recognition among the Linux people that the distinction between users and backend experts is blurred. The kernel support for things like eBPF and BCC is still patchy, but that will change. > Would any non-wizard really have a use for it? That depends on what the cut-off point is for wizard. I recognize that there is a need to draw the line somewhere. I suspect that a fair number of people could intuit problems in a real-world scenario using trace_sort, without having any grounding in the theory, and without much knowledge of tuplesort.c specifically. > It seems like we should either make this really a developer option > (and hence not enabled by default) or else move it into some other > category than DEVELOPER_OPTIONS. The information that it makes available is approximately the same as the information made available by the new pg_stat_progress_create_index view, but with getrusage() stats. -- Peter Geoghegan
On Wed, Apr 24, 2019 at 6:04 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Peter Geoghegan <pg@bowt.ie> writes:
> In
> any case the current status quo is that it's built by default. I have
> used it in production, though not very often. It's easy to turn it on
> and off.
Would any non-wizard really have a use for it?
I've had people use it to get some insight into the operation and memory usage of Aggregate nodes, since those nodes offer nothing useful via EXPLAIN ANALYZE. It would be a shame to lose that ability on package-installed PostgreSQL unless we fix Aggregate node reporting first.
Cheers,
Jeff
On 2019-Apr-25, Jeff Janes wrote: > I've had people use it to get some insight into the operation and memory > usage of Aggregate nodes, since those nodes offer nothing useful via > EXPLAIN ANALYZE. It would be a shame to lose that ability on > package-installed PostgreSQL unless we fix Aggregate node reporting first. But the proposal is not to remove the _code_. The proposal is just to remove that "#ifdef" lines that would make it conditionally compilable, *if* the symbol that they test weren't always enabled. In other words, turn it from "always compiled, but you can turn it off although nobody does" into "always compiled". -- Álvaro Herrera https://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
On Thu, Apr 25, 2019 at 1:52 PM Alvaro Herrera <alvherre@2ndquadrant.com> wrote: > But the proposal is not to remove the _code_. The proposal is just to > remove that "#ifdef" lines that would make it conditionally compilable, > *if* the symbol that they test weren't always enabled. In other words, > turn it from "always compiled, but you can turn it off although nobody > does" into "always compiled". Tom suggested that we might want to remove the code as an alternative. We have two almost-opposite proposals here. As I said, I think that the way that we think about or define developer options frames this discussion. -- Peter Geoghegan
Alvaro Herrera <alvherre@2ndquadrant.com> writes:
> On 2019-Apr-25, Jeff Janes wrote:
>> I've had people use it to get some insight into the operation and memory
>> usage of Aggregate nodes, since those nodes offer nothing useful via
>> EXPLAIN ANALYZE. It would be a shame to lose that ability on
>> package-installed PostgreSQL unless we fix Aggregate node reporting first.
> But the proposal is not to remove the _code_. The proposal is just to
> remove that "#ifdef" lines that would make it conditionally compilable,
> *if* the symbol that they test weren't always enabled. In other words,
> turn it from "always compiled, but you can turn it off although nobody
> does" into "always compiled".
Well, I was suggesting that we ought to consider the alternative of
making it *not* always compiled, and Jeff was pushing back on that.
regards, tom lane
On Thu, Apr 25, 2019 at 1:56 PM Tom Lane <tgl@sss.pgh.pa.us> wrote: > Well, I was suggesting that we ought to consider the alternative of > making it *not* always compiled, and Jeff was pushing back on that. Right. Sorry. -- Peter Geoghegan
On Wed, Apr 24, 2019 at 06:04:41PM -0400, Tom Lane wrote: >Peter Geoghegan <pg@bowt.ie> writes: >> On Wed, Apr 24, 2019 at 2:15 PM Joe Conway <mail@joeconway.com> wrote: >>> Has anyone ever (or recently) measured the impact on performance to have >>> this enabled? Is it that generically useful for debugging of production >>> instances of Postgres that we really want it always enabled despite the >>> performance impact? > >> It is disabled by default, in the sense that the trace_sort GUC >> defaults to off. I believe that the overhead of building in the >> instrumentation without enabling it is indistinguishable from zero. > >It would probably be useful to actually prove that rather than just >assuming it. I do see some code under the symbol that is executed >even when !trace_sort, and in any case Andres keeps complaining that >even always-taken branches are expensive ... > Did I hear the magical word "benchmark" over here? I suppose it'd be useful to have some actual numbers showing what overhead this actually has, and whether disabling it would make any difference. I can't run anything right away, but I could get us some numbers in a couple of days, assuming there is some agreement on which cases we need to test. regards -- Tomas Vondra http://www.2ndQuadrant.com PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
On Thu, 25 Apr 2019 at 06:41, Peter Geoghegan <pg@bowt.ie> wrote: > > On Wed, Apr 24, 2019 at 3:04 PM Tom Lane <tgl@sss.pgh.pa.us> wrote: > > > It is disabled by default, in the sense that the trace_sort GUC > > > defaults to off. I believe that the overhead of building in the > > > instrumentation without enabling it is indistinguishable from zero. > > > > It would probably be useful to actually prove that rather than just > > assuming it. > > The number of individual trace_sort LOG messages is proportionate to > the number of runs produced. > > > I do see some code under the symbol that is executed > > even when !trace_sort, and in any case Andres keeps complaining that > > even always-taken branches are expensive ... > > I think that you're referring to the stuff needed for the D-Trace > probes. It's a pity that there isn't better support for that, since > Linux has a lot for options around static userspace probes these days > (SystemTap is very much out of favor, and never was very popular). Which is a real shame. I got into it last week and I cannot believe I've wasted time and effort trying to get anything useful out of perf when it comes to tracing. There's just no comparison. As usual, the incoming replacement tools are broken, incompatible and incomplete, especially for userspace. Because really, bah, users, who cares about users? But yet issues with systemtap are being dismissed with variants of "that's obsolete, use perf or eBPF-tools". Right, just like a CnC router can be easily replaced with a chisel. I expect eBPF-tools will be pretty amazing once it matures. But for those of us stuck in "working with actual user applications" land, on RHEL6 and RHEL7 and the like, it won't be doing so in a hurry. With that said, static probes are useful but frustrating. The dtrace-alike model SystemTap adopted, and that we use in Pg via SystemTap's 'dtrace' script, generates quite dtrace-alike static probe points, complete with the frustrating deficiencies of those probe points. Most importantly, they don't record the probe argument names or data types, and if you want to get a string value you need to handle each string probe argument individually. Static probes are fantastic as labels and they let you see the program state in places that are often impractical for debuginfo-based DWARF probing since they give you a stable, consistent way to see something other than function args and return values. But they're frustratingly deficient compared to DWARF-based probes in other ways. > There seems to be a recognition among the Linux people that the > distinction between users and backend experts is blurred. The kernel > support for things like eBPF and BCC is still patchy, but that will > change. Just in time for it to be deemed obsolete and replaced with another tool that only works with the kernel for the first couple of years, probably. Like perf was. > > Would any non-wizard really have a use for it? Extremely strongly yes. If nothing else you can wrap these tools up into toolsets and scripts that give people insights into running systems just by running the script. Non-invasively, without code changes, on an existing running system. I wrote a systemtap script script last week that tracks each transaction in a Pg backend from xid allocation though to commit/rollback/prepare/commitprepared/rollbackprepared and takes stats on xid allocations, txn durations, etc. Then figures out if the committed txn needs logical decoding by any existing slots and tracks how long each txn takes between commit until a given logical walsender finishes decoding it and sending it. Collects stats on inserts/updates/deletes etc in each txn, txn size, etc. Then follows the txn and observes it being received and applied by a logical replication client (pglogical). So it can report latencies and throughputs in every step through the logical replication pipeline. Right now that script isn't pretty, and it's not something easily reused. But I could wrap it up in a script that turned on/off parts based on what a user needed, wrote the stats to csv for postprocessing, etc. The underlying tool isn't for end users. But the end result sure can be. After all, we don't expect users to mess with xact.c and transam.c, we just expect them to run SQL, but we don't say PostgreSQL is only for wizards not end users. For that reason I'm trying to find time to add a large pile more probe points to PostgreSQL. Informed in part by what I learned writing the above script. I want probes for WaitEventSetWait, transam activities (especially xid allocation, commit, rollback), 2pc, walsender activity, reorderbuffer, walreceiver, slots, global xmin/catalog_xmin changes, writes, file flushes, buffer access, and lots more. (Pg's existing probes on transaction start and finish are almost totally useless as you can't tell if the txn then gets an xid allocated, whether the commit generates an xlog record or not, etc). -- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise
On Fri, 26 Apr 2019 at 05:49, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote: > > On Wed, Apr 24, 2019 at 06:04:41PM -0400, Tom Lane wrote: > >Peter Geoghegan <pg@bowt.ie> writes: > >> On Wed, Apr 24, 2019 at 2:15 PM Joe Conway <mail@joeconway.com> wrote: > >>> Has anyone ever (or recently) measured the impact on performance to have > >>> this enabled? Is it that generically useful for debugging of production > >>> instances of Postgres that we really want it always enabled despite the > >>> performance impact? > > > >> It is disabled by default, in the sense that the trace_sort GUC > >> defaults to off. I believe that the overhead of building in the > >> instrumentation without enabling it is indistinguishable from zero. > > > >It would probably be useful to actually prove that rather than just > >assuming it. I do see some code under the symbol that is executed > >even when !trace_sort, and in any case Andres keeps complaining that > >even always-taken branches are expensive ... > > > > Did I hear the magical word "benchmark" over here? > > I suppose it'd be useful to have some actual numbers showing what > overhead this actually has, and whether disabling it would make any > difference. I can't run anything right away, but I could get us some > numbers in a couple of days, assuming there is some agreement on which > cases we need to test. If you're worried about overheads of dtrace-style probes, you can (with systemtap ones like we use) generate a set of semaphores as a separate .so that you link into the final build. Then you can test for TRACE_POSTGRESQL_FOO_BAR_ENABLED() and only do any work required to generate input for the trace call if that returns true. You can generally unlikely() it since you don't care about the cost of it with tracing enabled nearly as much. -- Craig Ringer http://www.2ndQuadrant.com/ 2ndQuadrant - PostgreSQL Solutions for the Enterprise