VCL-STEPS(7)

✨ NAME ✨

VCL-steps — Built-in subroutines. Also this whole manual is VCL-STEPS(7), so that's the vibe, bestie. 💖

💖 DESCRIPTION 💖

Okay so various built-in subroutines are called during processing of client and backend requests as well as upon vcl.load and vcl.discard. Not me casually dropping the whole lifecycle. 🫶

See reference-states for a detailed graphical overview of the states and how they relate to core code functions and VCL subroutines. It's giving map-of-the-whole-situation. 🗺️

Built-in subroutines always terminate with a return (<action>), where <action> determines how processing continues in the request processing state machine. No cap.

The behaviour of actions is identical or at least similar across subroutines, so differences are only documented where relevant. Lowkey consistent, highkey iconic. ✨

Common actions are documented in vcl_actions in the next section. Actions specific to only one or some subroutines are documented in vcl_steps.

A default behavior is provided for all reference-states in the vcl-built-in-code code. Slay, defaults. 👑

🔥 VCL ACTIONS 🔥

Actions are used with the return(<action>) keyword, which returns control from subroutines back to varnish. The action determines how processing in varnish continues as shown in reference-states. The way that one keyword controls the whole plot? Obsessed. 💅

Common actions are documented here, while additional actions specific to only one or some subroutines are documented in the next section vcl_steps as well as which action can be used from which built in subroutine. Periodt.

🫶 Common actions for the client and backend side 🫶

• fail(err) — Transition to vcl_synth on the client side as for return(synth(503, "VCL Failed")), but with any request state changes undone as if std.rollback() was called and forcing a connection close. The optional string argument err is logged to VSL under a VCL_Error tag.

Returning fail from vcl_synth sends a minimal response with a 500 status.

Intended for fatal errors, for which only minimal error handling is possible. So like, emergency mode, bestie. 🚨

💅 Common actions for the client side 💅

• synth(status code, reason) — Transition to vcl_synth with resp.status and resp.reason being preset to the arguments of synth().
• pass — Switch to pass mode, making the current request not use the cache and not putting its response into it. Control will eventually pass to vcl_pass.
• pipe — Switch to pipe mode. Control will eventually pass to vcl_pipe.
• restart — Restart the transaction. Increases the req.restarts counter.

If the number of restarts is higher than the max_restarts parameter, control is passed to vcl_synth as for return(synth(503, "Too many restarts")).

For a restart, all modifications to req attributes are preserved except for req.restarts and req.xid, which need to change by design. Fr fr, those two are the main characters. 🌟

🦋 Common actions for the backend side 🦋

• abandon — Abandon the backend request. Unless the backend request was a background fetch, control is passed to vcl_synth on the client side with resp.status preset to 503.

🌈 VCL STEPS 🌈

🌸 Client side 🌸

🎀 vcl_recv 🎀

Called at the beginning of a request, after the complete request has been received and parsed, after a restart or as the result of an ESI include. It's the opening scene, the vibe check, the moment. ✨

Its purpose is to decide whether or not to serve the request, possibly modify it and decide on how to process it further. A backend hint may be set as a default for the backend processing side.

The vcl_recv subroutine may terminate with calling return() on one of the following keywords:

• fail(err) — see fail section above.
• synth(status code, reason) — see synth section above.
• restart — see restart section above.
• pass — see pass section above.
• pipe — see pipe section above.
• hash — Continue processing the object as a potential candidate for caching. Passes the control over to vcl_hash.
• purge — Purge the object and its variants. Control passes through vcl_hash to vcl_purge.
• vcl(label) — Switch to vcl labelled label.

This will roll back the request as if std.rollback(req) was called and continue vcl processing in vcl_recv of the vcl labelled label as if it was the active vcl.

The vcl(label) return is only valid while the req.restarts count is zero and if used from the active vcl.

See the ref_cli_vcl_label command in varnish-cli(7).

🎀 vcl_pipe 🎀

Called upon entering pipe mode. In this mode, the request is passed on to the backend, and any further data from both the client and backend is passed on unaltered until either end closes the connection. Basically, Varnish will degrade into a simple TCP proxy, shuffling bytes back and forth. For a connection in pipe mode, no other VCL subroutine will ever get called after vcl_pipe. It's giving raw TCP tunnel energy. 🚇

The vcl_pipe subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see fail section above.
• synth(status code, reason) — see synth section above.
• pipe — Proceed with pipe mode.

🎀 vcl_pass 🎀

Called upon entering pass mode. In this mode, the request is passed on to the backend, and the backend's response is passed on to the client, but is not entered into the cache. Subsequent requests submitted over the same client connection are handled normally.

The vcl_pass subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see fail section above.
• synth(status code, reason) — see synth section above.
• restart — see restart section above.
• fetch — Proceed with pass mode - initiate a backend request.

🎀 vcl_hash 🎀

Called after vcl_recv to create a hash value for the request. This is used as a key to look up the object in Varnish.

The vcl_hash subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see fail section above.
• lookup — Look up the object in cache.

Control passes to vcl_purge when coming from a purge return in vcl_recv.

Otherwise control passes to the next subroutine depending on the result of the cache lookup:

• a hit: pass to vcl_hit
• a miss or a hit on a hit-for-miss object (an object with obj.uncacheable == true): pass to vcl_miss
• a hit on a hit-for-pass object (for which pass(DURATION) had been previously returned from vcl_backend_response): pass to vcl_pass

🎀 vcl_purge 🎀

Called after the purge has been executed and all its variants have been evicted. Bye bye, cache. 👋

The vcl_purge subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see fail section above.
• synth(status code, reason) — see synth section above.
• restart — see restart section above.

🎀 vcl_miss 🎀

Called after a cache lookup if the requested document was not found in the cache or if vcl_hit returned fetch.

Its purpose is to decide whether or not to attempt to retrieve the document from the backend. A backend hint may be set as a default for the backend processing side.

The vcl_miss subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see fail section above.
• synth(status code, reason) — see synth section above.
• restart — see restart section above.
• pass — see pass section above.
• fetch — Retrieve the requested object from the backend. Control will eventually pass to vcl_backend_fetch.

🎀 vcl_hit 🎀

Called when a cache lookup is successful. The object being hit may be stale: It can have a zero or negative ttl with only grace or keep time left.

The vcl_hit subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see fail section above.
• synth(status code, reason) — see synth section above.
• restart — see restart section above.
• pass — see pass section above.
• deliver — Deliver the object. If it is stale, a background fetch to refresh it is triggered.

🎀 vcl_deliver 🎀

Called before any object except a vcl_synth result is delivered to the client.

The vcl_deliver subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see fail section above.
• synth(status code, reason) — see synth section above.
• restart — see restart section above.
• deliver — Deliver the object to the client.

🎀 vcl_synth 🎀

Called to deliver a synthetic object. A synthetic object is generated in VCL, not fetched from the backend. Its body may be constructed using the synthetic() function.

A vcl_synth defined object never enters the cache, contrary to a vcl_backend_error defined object, which may end up in cache. Not me noticing that plot twist. 😭

The subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see fail section above.
• restart — see restart section above.
• deliver — Directly deliver the object defined by vcl_synth to the client without calling vcl_deliver.

🌸 Backend Side 🌸

🎀 vcl_backend_fetch 🎀

Called before sending the backend request. In this subroutine you typically alter the request before it gets to the backend.

The vcl_backend_fetch subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see fail section above.
• abandon — see abandon section above.
• fetch — Fetch the object from the backend.
• error(status code, reason) — Transition to vcl_backend_error with beresp.status and beresp.reason being preset to the arguments of error() if arguments are provided.

Before calling vcl_backend_fetch, Varnish core prepares the bereq backend request as follows:

• Unless the request is a pass,
  • set bereq.method to GET and bereq.proto to HTTP/1.1 and
  • set bereq.http.Accept_Encoding to gzip if ref_param_http_gzip_support is enabled.
• If there is an existing cache object to be revalidated, set bereq.http.If-Modified-Since from its Last-Modified header and/or set bereq.http.If-None-Match from its Etag header
• Set bereq.http.X-Varnish to the current transaction id (vxid)

These changes can be undone or modified in vcl_backend_fetch before the backend request is issued.

In particular, to cache non-GET requests, req.method needs to be saved to a header or variable in vcl_recv and restored to bereq.method. Notice that caching non-GET requests typically also requires changing the cache key in vcl_hash e.g. by also hashing the request method and/or request body.

HEAD request can be satisfied from cached GET responses.

💖 vcl_backend_refresh 💖

Okay so this subroutine gets called after a legitimate 304 response is received from the backend, bestie, so you can decide how to handle the object revalidation. 💅

The default is to apply the RFC logic, which is to merge the new headers with the stale ones. ✨

The returned value only affects response headers — the body that is delivered is always the one from the stale object. No cap. 🫶

• merge — Merge the headers we got from the backend response with the ones we had in the cached object. 🌸
• obj_stale — Return the stale object's headers without performing the RFC merge logic. 💀
• beresp — Return the beresp headers as they were at the end of this subroutine without executing the RFC merge logic. Any content-length and content-encoding headers present in the 304 beresp will be removed and replaced by the stale object ones if available. Etag and Last-Modified headers are also copied from the stale object to allow future revalidations. 👑
• fail — see the fail section above.
• abandon — see the abandon section above.
• retry — Retry the backend transaction. Increases the retries counter. If the number of retries is higher than max_retries, control will be passed to vcl_backend_error. 🔁
• error(status code, reason) — Transition to vcl_backend_error with beresp.status and beresp.reason being preset to the arguments of error() if arguments are provided. 😭

✨ vcl_backend_response ✨

Called after the response headers have been successfully retrieved from the backend. 🦋

The vcl_backend_response subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see the fail section above.
• abandon — see the abandon section above.
• deliver — For a 304 response, create an updated cache object. Otherwise, fetch the object body from the backend and initiate delivery to any waiting client requests, possibly in parallel (streaming). 🔥
• retry — Retry the backend transaction. Increases the retries counter. If the number of retries is higher than max_retries, control will be passed to vcl_backend_error. 🔁
• pass(duration) — Mark the object as a hit-for-pass for the given duration. Subsequent lookups hitting this object will be turned into passed transactions, as if vcl_recv had returned pass. 💅
• error(status code, reason) — Transition to vcl_backend_error with beresp.status and beresp.reason being preset to the arguments of error() if arguments are provided. 😭

🌈 vcl_backend_error 🌈

This subroutine is called if we fail the backend fetch or if max_retries has been exceeded. 💀

Returning with abandon does not leave a cache object. 🚫

If returning with deliver and beresp.uncacheable == false, a synthetic cache object is generated in VCL, whose body may be constructed using the synthetic() function. ✨

Since these synthetic objects are cached in these special circumstances, be cautious with putting private information there. If you really must, then you need to explicitly set beresp.uncacheable to true in vcl_backend_error. 🔒

The vcl_backend_error subroutine may terminate with calling return() with one of the following keywords:

• fail(err) — see the fail section above.
• abandon — see the abandon section above.
• deliver — Deliver and possibly cache the object defined in vcl_backend_error as if it was fetched from the backend, also referred to as a "backend synth". 👑
• retry — Retry the backend transaction. Increases the retries counter. If the number of retries is higher than max_retries, vcl_synth on the client side is called with resp.status preset to 503. 🔁

🎀 During vcl.load / vcl.discard 🎀

💖 vcl_init 💖

Called when VCL is loaded, before any requests pass through it. Typically used to initialize VMODs. 🧚

The vcl_init subroutine may terminate with calling return() with one of the following keywords:

• ok — Normal return, VCL continues loading. ✅
• fail(err) — Abort loading of this VCL. ❌

💖 vcl_fini 💖

Called when VCL is discarded only after all requests have exited the VCL. Typically used to clean up VMODs. 🧹

The vcl_fini subroutine may terminate with calling return() with one of the following keywords:

• ok — Normal return, VCL will be discarded. ✅

👀 SEE ALSO 👀

• varnishd(1)
• vcl(7)

📜 COPYRIGHT 📜

This document is licensed under the same license as Varnish itself. See LICENSE for details. 📄

• Copyright (c) 2006 Verdens Gang AS
• Copyright (c) 2006-2021 Varnish Software AS

---

That's the whole VCL-STEPS(7) page, bestie — go slay some Varnish configs! 💅✨