Comencemos con un ejemplo y luego veámoslo en detalle.

Creamos una función asíncrona `lifespan()` con `yield` así:

{* ../../docs_src/events/tutorial003_py39.py hl[16,19] *}

Nós criamos uma função assíncrona `lifespan()` com `yield` assim:

{* ../../docs_src/events/tutorial003_py39.py hl[16,19] *}

Aqui estamos simulando a operação de *inicialização* custosa de carregar o modelo, colocando a (falsa) função do modelo no dicionário com modelos de machine learning antes do `yield`. Esse código será executado **antes** de a aplicação **começar a receber requisições**, durante a *inicialização*.

 * could in principle cause problems for some users. Still, we expect that the benefits of the
 * nullness annotations in particular will outweigh the costs. (And it's worth noting that we have
 * released multiple ListenableFuture.class files that are not byte-for-byte compatible even from
 * the beginning, thanks to using different `-source -target` values for compiling our `-jre` and

 * could in principle cause problems for some users. Still, we expect that the benefits of the
 * nullness annotations in particular will outweigh the costs. (And it's worth noting that we have
 * released multiple ListenableFuture.class files that are not byte-for-byte compatible even from
 * the beginning, thanks to using different `-source -target` values for compiling our `-jre` and

 * could in principle cause problems for some users. Still, we expect that the benefits of the
 * nullness annotations in particular will outweigh the costs. (And it's worth noting that we have
 * released multiple ListenableFuture.class files that are not byte-for-byte compatible even from
 * the beginning, thanks to using different `-source -target` values for compiling our `-jre` and