Transient storage killing memory

[Niall Murphy]

On Tue, 29 Nov 2016 17:32:17 +0100
Dridi Boukelmoune <dridi at varni.sh> wrote:

> > I took a look into ttl < 10s objects and saw that they are requests
> > we intentionally apply either "max-age=0, private" or "no-cache,
> > private" to. However their storage field is still "malloc
> > Transient", and transient storage usage only appears to be going
> > up.  
> 
> Transient storage is used for short-lived object, but also for pass'ed
> transactions. The storage is freed when the response is consumed by
> the client.
> 
> Which counters appears to only go up?
> 
SMA.s0.c_req                            210375         2.65 Allocator requests
SMA.s0.c_fail                                0         0.00 Allocator failures
SMA.s0.c_bytes                       976439262     12317.58 Bytes allocated
SMA.s0.c_freed                       888098949     11203.19 Bytes freed
SMA.s0.g_alloc                           22425          .   Allocations outstanding
SMA.s0.g_bytes                        88340313          .   Bytes outstanding
SMA.s0.g_space                       698091687          .   Bytes available
SMA.Transient.c_req                     622804         7.86 Allocator requests
SMA.Transient.c_fail                         0         0.00 Allocator failures
SMA.Transient.c_bytes               2780556626     35076.15 Bytes allocated
SMA.Transient.c_freed               2224419585     28060.60 Bytes freed
SMA.Transient.g_alloc                   337153          .   Allocations outstanding
SMA.Transient.g_bytes                556137041          .   Bytes outstanding
SMA.Transient.g_space                        0          .   Bytes available

Transient I thought but it appears it _is_ being freed at least, but
it's unusually large compared to s0 when you consider only 7% of
requests have a ttl < shortlived wouldn't you say?

Synthetic error responses are low, but hit-pass (return(miss) in v5)
are high:

MAIN.client_req           597833         7.52 Good client requests received
MAIN.cache_hit            289103         3.64 Cache hits
MAIN.cache_hitpass        284475         3.58 Cache hits for pass
MAIN.cache_miss           307426         3.87 Cache misses

But I only tell varnish to miss on the above 7%.