Re: [PATCH] memtx: add yields during index build

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

On Wed, May 22, 2019 at 07:11:42PM +0300, Serge Petrenko wrote:
> Memtx index build used to stall event loop for all the build period.
> Add occasional yields so that the loop is not blocked for too long.
> 
> Closes #3976
> ---
> https://github.com/tarantool/tarantool/issues/3976
> https://github.com/tarantool/tarantool/tree/sp/gh-3976-background-index-build
> 
>  src/box/memtx_space.c                         | 13 +++++
>  test/box/memtx_background_index_build.result  | 55 +++++++++++++++++++
>  .../box/memtx_background_index_build.test.lua | 32 +++++++++++
>  test/box/suite.ini                            |  2 +-
>  4 files changed, 101 insertions(+), 1 deletion(-)
>  create mode 100644 test/box/memtx_background_index_build.result
>  create mode 100644 test/box/memtx_background_index_build.test.lua
> 
> diff --git a/src/box/memtx_space.c b/src/box/memtx_space.c
> index 5ddb4f7ee..b90e2707e 100644
> --- a/src/box/memtx_space.c
> +++ b/src/box/memtx_space.c
> @@ -874,6 +874,15 @@ static int
>  memtx_space_build_index(struct space *src_space, struct index *new_index,
>  			struct tuple_format *new_format)
>  {
> +	/*
> +	 * Yield every 1K tuples.
> +	 * In debug mode yield more often for testing purposes.
> +	 */
> +#ifdef NDEBUG
> +	enum { YIELD_LOOPS = 1000 };
> +#else
> +	enum { YIELD_LOOPS = 10 };
> +#endif
>  	/**
>  	 * If it's a secondary key, and we're not building them
>  	 * yet (i.e. it's snapshot recovery for memtx), do nothing.
> @@ -909,6 +918,7 @@ memtx_space_build_index(struct space *src_space, struct index *new_index,
>  	/* Build the new index. */
>  	int rc;
>  	struct tuple *tuple;
> +	size_t count = 0;
>  	while ((rc = iterator_next(it, &tuple)) == 0 && tuple != NULL) {
>  		/*
>  		 * Check that the tuple is OK according to the
> @@ -933,6 +943,9 @@ memtx_space_build_index(struct space *src_space, struct index *new_index,
>  		 */
>  		if (new_index->def->iid == 0)
>  			tuple_ref(tuple);
> +		if (++count % YIELD_LOOPS == 0) {
> +			fiber_sleep(0);
> +		}

This isn't enough as tuples may be inserted into / deleted from the
space while DDL is in progress. We need to propagate these changes
to the new index using on_replace trigger. Take a look at how vinyl
handles it.