Re: [tarantool-patches] [PATCH] Do not enable commit if read_only = true

[Vladimir Davydov <vdavydov.dev@gmail.com>]

On Wed, Feb 27, 2019 at 10:36:00AM +0300, Georgy Kirichenko wrote:
> Disable commit if server is in read only mode.

The commit message is very poor. Please elaborate why this is important.

> 
> Closes: #4016
> ---
> Issue: https://github.com/tarantool/tarantool/issues/4016
> Branch: https://github.com/tarantool/tarantool/tree/g.kirichenko/gh-4016-readonly-commit
>  src/box/box.cc         |  2 +-
>  src/box/box.h          |  3 +++
>  src/box/txn.c          |  6 ++++++
>  test/box/misc.result   | 19 +++++++++++++++++++
>  test/box/misc.test.lua |  8 ++++++++
>  5 files changed, 37 insertions(+), 1 deletion(-)
> 
> diff --git a/src/box/box.cc b/src/box/box.cc
> index 73d94f79b..ec196bcc0 100644
> --- a/src/box/box.cc
> +++ b/src/box/box.cc
> @@ -138,7 +138,7 @@ static struct fiber_pool tx_fiber_pool;
>   */
>  static struct cbus_endpoint tx_prio_endpoint;
>  
> -static int
> +int
>  box_check_writable(void)
>  {
>  	/* box is only writable if box.cfg.read_only == false and */
> diff --git a/src/box/box.h b/src/box/box.h
> index 9f5b3acbd..d9e403d7c 100644
> --- a/src/box/box.h
> +++ b/src/box/box.h
> @@ -101,6 +101,9 @@ box_set_ro(bool ro);
>  bool
>  box_is_ro(void);
>  
> +int
> +box_check_writable(void);
> +
>  /**
>   * Wait until the instance switches to a desired mode.
>   * \param ro wait read-only if set or read-write if unset
> diff --git a/src/box/txn.c b/src/box/txn.c
> index d55d5b93c..769a57a5a 100644
> --- a/src/box/txn.c
> +++ b/src/box/txn.c
> @@ -34,6 +34,7 @@
>  #include "journal.h"
>  #include <fiber.h>
>  #include "xrow.h"
> +#include "box.h"

Ouch. Can we avoid introducing this dependency?

>  
>  double too_long_threshold;
>  
> @@ -448,6 +449,11 @@ box_txn_commit()
>  	*/
>  	if (! txn)
>  		return 0;
> +	/*
> +	 * Check that tarantool didn't switch to ro.
> +	 */
> +	if (box_check_writable() != 0)
> +		return -1;

What about temporary and local spaces? We don't want this check to fail
transactions for those. Please fix and add a corresponding test case.

Also, may be it's worth moving the ro check completely to txn_commit?
IMO it looks weird that we check it both when processing a request and
when committing a transaction.

An alternative approach would be setting a trigger on yield and checking
that we are still rw on resume, aborting transactions if we are not.
This would remove the check on txn_commit and probably allow us to
eliminate box.h dependency. Please check it out.

Also, please try to implement a test that checks this for vinyl +
replication. After all, this problem is only relevant to vinyl.

>  	if (txn->in_sub_stmt) {
>  		diag_set(ClientError, ER_COMMIT_IN_SUB_STMT);
>  		return -1;