1 gRPC Connectivity Semantics and API 2 =================================== 3 4 This document describes the connectivity semantics for gRPC channels and the 5 corresponding impact on RPCs. We then discuss an API. 6 7 States of Connectivity 8 ---------------------- 9 10 gRPC Channels provide the abstraction over which clients can communicate with 11 servers.The client-side channel object can be constructed using little more 12 than a DNS name. Channels encapsulate a range of functionality including name 13 resolution, establishing a TCP connection (with retries and backoff) and TLS 14 handshakes. Channels can also handle errors on established connections and 15 reconnect, or in the case of HTTP/2 GO_AWAY, re-resolve the name and reconnect. 16 17 To hide the details of all this activity from the user of the gRPC API (i.e., 18 application code) while exposing meaningful information about the state of a 19 channel, we use a state machine with five states, defined below: 20 21 CONNECTING: The channel is trying to establish a connection and is waiting to 22 make progress on one of the steps involved in name resolution, TCP connection 23 establishment or TLS handshake. This may be used as the initial state for channels upon 24 creation. 25 26 READY: The channel has successfully established a connection all the way 27 through TLS handshake (or equivalent) and all subsequent attempt to communicate 28 have succeeded (or are pending without any known failure ). 29 30 TRANSIENT_FAILURE: There has been some transient failure (such as a TCP 3-way 31 handshake timing out or a socket error). Channels in this state will eventually 32 switch to the CONNECTING state and try to establish a connection again. Since 33 retries are done with exponential backoff, channels that fail to connect will 34 start out spending very little time in this state but as the attempts fail 35 repeatedly, the channel will spend increasingly large amounts of time in this 36 state. For many non-fatal failures (e.g., TCP connection attempts timing out 37 because the server is not yet available), the channel may spend increasingly 38 large amounts of time in this state. 39 40 IDLE: This is the state where the channel is not even trying to create a 41 connection because of a lack of new or pending RPCs. New RPCs MAY be created 42 in this state. Any attempt to start an RPC on the channel will push the channel 43 out of this state to connecting. When there has been no RPC activity on a channel 44 for a specified IDLE_TIMEOUT, i.e., no new or pending (active) RPCs for this 45 period, channels that are READY or CONNECTING switch to IDLE. Additionaly, 46 channels that receive a GOAWAY when there are no active or pending RPCs should 47 also switch to IDLE to avoid connection overload at servers that are attempting 48 to shed connections. We will use a default IDLE_TIMEOUT of 300 seconds (5 minutes). 49 50 SHUTDOWN: This channel has started shutting down. Any new RPCs should fail 51 immediately. Pending RPCs may continue running till the application cancels them. 52 Channels may enter this state either because the application explicitly requested 53 a shutdown or if a non-recoverable error has happened during attempts to connect 54 communicate . (As of 6/12/2015, there are no known errors (while connecting or 55 communicating) that are classified as non-recoverable) 56 Channels that enter this state never leave this state. 57 58 The following table lists the legal transitions from one state to another and 59 corresponding reasons. Empty cells denote disallowed transitions. 60 61 <table style='border: 1px solid black'> 62 <tr> 63 <th>From/To</th> 64 <th>CONNECTING</th> 65 <th>READY</th> 66 <th>TRANSIENT_FAILURE</th> 67 <th>IDLE</th> 68 <th>SHUTDOWN</th> 69 </tr> 70 <tr> 71 <th>CONNECTING</th> 72 <td>Incremental progress during connection establishment</td> 73 <td>All steps needed to establish a connection succeeded</td> 74 <td>Any failure in any of the steps needed to establish connection</td> 75 <td>No RPC activity on channel for IDLE_TIMEOUT</td> 76 <td>Shutdown triggered by application.</td> 77 </tr> 78 <tr> 79 <th>READY</th> 80 <td></td> 81 <td>Incremental successful communication on established channel.</td> 82 <td>Any failure encountered while expecting successful communication on 83 established channel.</td> 84 <td>No RPC activity on channel for IDLE_TIMEOUT <br>OR<br>upon receiving a GOAWAY while there are no pending RPCs.</td> 85 <td>Shutdown triggered by application.</td> 86 </tr> 87 <tr> 88 <th>TRANSIENT_FAILURE</th> 89 <td>Wait time required to implement (exponential) backoff is over.</td> 90 <td></td> 91 <td></td> 92 <td></td> 93 <td>Shutdown triggered by application.</td> 94 </tr> 95 <tr> 96 <th>IDLE</th> 97 <td>Any new RPC activity on the channel</td> 98 <td></td> 99 <td></td> 100 <td></td> 101 <td>Shutdown triggered by application.</td> 102 </tr> 103 <tr> 104 <th>SHUTDOWN</th> 105 <td></td> 106 <td></td> 107 <td></td> 108 <td></td> 109 <td></td> 110 </tr> 111 </table> 112 113 114 Channel State API 115 ----------------- 116 117 All gRPC libraries will expose a channel-level API method to poll the current 118 state of a channel. In C++, this method is called GetState and returns an enum 119 for one of the five legal states. It also accepts a boolean `try_to_connect` to 120 transition to CONNECTING if the channel is currently IDLE. The boolean should 121 act as if an RPC occurred, so it should also reset IDLE_TIMEOUT. 122 123 ```cpp 124 grpc_connectivity_state GetState(bool try_to_connect); 125 ``` 126 127 All libraries should also expose an API that enables the application (user of 128 the gRPC API) to be notified when the channel state changes. Since state 129 changes can be rapid and race with any such notification, the notification 130 should just inform the user that some state change has happened, leaving it to 131 the user to poll the channel for the current state. 132 133 The synchronous version of this API is: 134 135 ```cpp 136 bool WaitForStateChange(grpc_connectivity_state source_state, gpr_timespec deadline); 137 ``` 138 139 which returns `true` when the state is something other than the 140 `source_state` and `false` if the deadline expires. Asynchronous- and futures-based 141 APIs should have a corresponding method that allows the application to be 142 notified when the state of a channel changes. 143 144 Note that a notification is delivered every time there is a transition from any 145 state to any *other* state. On the other hand the rules for legal state 146 transition, require a transition from CONNECTING to TRANSIENT_FAILURE and back 147 to CONNECTING for every recoverable failure, even if the corresponding 148 exponential backoff requires no wait before retry. The combined effect is that 149 the application may receive state change notifications that appear spurious. 150 e.g., an application waiting for state changes on a channel that is CONNECTING 151 may receive a state change notification but find the channel in the same 152 CONNECTING state on polling for current state because the channel may have 153 spent infinitesimally small amount of time in the TRANSIENT_FAILURE state. 154
