I have 2 NextJS internal servers/applications that is reverse-proxied by Apache 2.4. Let's have the 2 internal servers as http://internal:3000/foo
and http://internal:3001/foo
and the external URL as http://external/foo
.
When the NextJS is accessed on the base path (i.e. http://internal:3000/foo
), it will be redirected to http://internal:3000/foo/bar/baz
on HTTP Code 308
. So over the reverse-proxy, I expected the same, that the redirection will happen from http://external/foo
to http://external/foo/bar/baz
.
The config in next.config.js
is as follow
...
module.exports = {
...
async redirects() {
return [
{
source: '/',
destination: '/bar/baz',
permanent: true
}
]
},
basePath: 'foo'
}
What happen is that this redirection works perfectly when I tried reverse-proxy to only 1 NextJS application without load balancing, e.g. I only reverse proxy to http://internal:3000/foo
.
The config that I used is as follow
<Location "/foo">
ProxyPass "http://localhost:3000/foo"
ProxyPassReverse "http://localhost:3000/foo"
</Location>
But the redirection does not work when I tried reverse-proxy to 2 NextJS application on load balancing.
The config that I used is as follow
<Proxy "balancer://example">
BalancerMember "http://localhost:3000/foo"
BalancerMember "http://localhost:3001/foo"
</Proxy>
<Location "/foo">
ProxyPass "balancer://example"
ProxyPassReverse "balancer://example"
</Location>
What happen instead is that it will keep redirecting from http://external/foo
to http://external/foo
i.e. infinite redirect that result in TOO_MANY_REDIRECT
.
It baffles me that the redirection works in non load-balancing scenario but failed when using load-balancing. Any ideas what is actually happening? Is there response header being written that I am unaware of when using proxy load-balancing? Thanks!
Update/progress (1):
I suspect there is rewriting happening in mod_proxy_balancer.c
in following section
access_status = rewrite_url(r, *worker, url);
/* Add the session route to request notes if present */
if (route) {
apr_table_setn(r->notes, "session-sticky", sticky);
apr_table_setn(r->notes, "session-route", route);
/* Add session info to env. */
apr_table_setn(r->subprocess_env,
"BALANCER_SESSION_STICKY", sticky);
apr_table_setn(r->subprocess_env,
"BALANCER_SESSION_ROUTE", route);
}
ap_log_rerror(APLOG_MARK, APLOG_DEBUG, 0, r, APLOGNO(01172)
"%s: worker (%s) rewritten to %s",
(*balancer)->s->name, (*worker)->s->name, *url);
return access_status;
which can be found in https://github.com/apache/httpd/blob/317108ee6e84ae47bd0f6121e3a64074c5d68c7b/modules/proxy/mod_proxy_balancer.c#L631-L647
Update/progress (2):
I turned on mod_dumpio
to log all the incoming and outgoing traffic into and from apache, and have confirmed that there is indeed a rewriting happening.
The rewriting is happening as follow:
GET /foo
, which is the original request sent to the external server.GET /foo/
, which is the request sent to the internal server. Notice there is rewriting of adding slash at the end.Location /foo
, which is the internal server's redirection response location. This is an intended location because in internal server, GET /foo/
will be redirected to Location /foo
while GET /foo
will be redirected to Location /foo/bar/baz
. Under normal circumstances the redirection will be handled by the internal server, meaning that GET /foo/
will result in redirection resulting in GET /foo
, eventually yielding Location /foo/bar/baz
but this does not happen in reverse-proxy.Location /foo
, which is the external server's redirection response location. Because (4) and (1) is the same URL, therefore it will create a redirection loop.With the rewriting confirmed, now I am looking whether there is way to solve this behavior.
I finally managed to fix the issue, and pertaining to the 2nd update above, I managed to find where the "rewriting" happen which helped me fix the issue.
tl;dr
The fix is basically moving the path from BalancerMember
to the balancer itself, i.e.
<Proxy "balancer://example/foo"> // used to be "balancer://example"
BalancerMember "http://localhost:3000" // used to be "http://localhost:3000/foo"
BalancerMember "http://localhost:3001" // used to be "http://localhost:3001/foo"
</Proxy>
// change to point to the balancer accordingly
<Location "/foo">
ProxyPass "balancer://example/foo"
ProxyPassReverse "balancer://example/foo"
</Location>
As for my finding, the "rewriting" is not so much of rewriting but is actually URL canonicalisation performed by Apache, in mod_proxy_balancer
or mod_proxy_http
modules (or depends on your scheme). example of canonicalisation source code in mod_proxy_balancer
Changing balancer://example -> balancer://example/foo
makes its URL scheme structure wise same to http://localhost:3000/foo
, which when undergoing canonicalisation will not yield trailing slash at the end, which is caused by balancer://example
not having path right after the "host", and therefore with the said change the behaviour for reverse-proxy using either load balancer or not will finally be the same.