Versioning in the API is supported by so-called TL layers.
The need to add a new object constructor or to add/remove a field in a constructor creates a backwards compatibility problem for previous versions of API clients. After all, simply changing a constructor in a schema also changes its number. To address this problem, each schema update is separated into a layer.
A layer is a collection of updated methods or constructors in a TL schema. Each layer is numbered with sequentially increasing numbers starting with 2. The first layer is the base layer -- the TL schema without any changes.
There is a helper method to let the API know that a client supports the Layer layer:
invokeWithLayer#da9b0d0d {X:Type} layer:int query:!X = X;The helper method invokeWithLayer can be used only together with initConnection: the present layer will be saved with all other parameters of the client and any future requests will be using this saved value. See more below.
It is possible to save information about the current client on the server in conjunction with an authorization key. This may help eliminate client-side problems with certain releases on certain devices or with certain localizations, as well as eliminate the need for sending layer information in each request.
The helper method initConnection accepts client parameters. This method must be called when first calling the API after the application has restarted or in case the value of one of the parameters could have changed.
initConnection must also be called after each auth.bindTempAuthKey.
When calling this method, the current layer used by the client is also saved (the layer in which initConnection was wrapped is used). After a successful call to initConnection it is no longer necessary to wrap each API call in invokeWithLayerN.
invokeWithoutUpdates#bf9459b7 {X:Type} query:!X = X;invokeWithoutUpdates can be used to invoke a request without subscribing the used connection for updates (this is enabled by default for file queries).
By default, the server processes parallel requests in arbitrary order. Two helper methods exist for cases when the client needs certain requests to be processed in a certain order and intends to send a new request before the previous one is completed.
These methods allow reducing the latency for calls which require strict ordering, since the client doesn't have to wait for the result of the previous method call before sending the next one in the queue, and can just send them all together, (for example) each wrapped in an invokeAfterMsg with the msg_id of the previous request.
invokeAfterMsg#cb9f372d {X:Type} msg_id:long query:!X = X;
invokeAfterMsgs#3dc4b4f0 {X:Type} msg_ids:Vector
They may be used, for example, when a client attempts to send messages that accumulated while waiting for the Internet connection to be restored for a long time. In this case, the 32-bit number 0xcb9f372d must be added before the method number in each request, followed by a 64-bit message identifier, msg_id, which contains the previous request in the queue.
The second method is similar, except it takes several messages that must be successfully processed before the current one.
If the waiting period exceeds 0.5 seconds (this value may change in the future) and no result has appeared, the method will return the MSG_WAIT_TIMEOUT error: handle this error by resending the request, still wrapped in the same invokeAfterMsg/invokeAfterMsgs constructor with the same id/ids.
If any of the previous queries mentioned in msg_ids or msg_id fails (i.e. an RPC error is emitted by a query, including FLOOD_WAIT_ errors), a MSG_WAIT_FAILED error will be returned for the current request: the simplest way to handle it is to simply enforce local synchronization, by waiting for a response from all previous msg_ids/msg_id before resending the request.
If and only if any of the previous requests also failed with MSG_WAIT_FAILED/MSG_WAIT_TIMEOUT errors and require resending, wrap the current request in another invokeAfterMsg/invokeAfterMsgs constructor with the new IDs of the previous requests, resent along with the current one.
To clarify, assume the following sequence of queries:
messages.sendMessage message=ainvokeAfterMsg msg_id=1 (messages.sendMessage message=b)invokeAfterMsg msg_id=2 (messages.sendMessage message=c)If the first messages.sendMessage query with msg_id=1 fails, both queries with msg_id=2 and msg_id=3 will fail with a MSG_WAIT_FAILED and will have to be resent as follows.
To recover the call queue, send the following new sequence of queries:
messages.sendMessage message=b (resending the old query with msg_id=2)invokeAfterMsg msg_id=4 (messages.sendMessage message=c) (resending the old query with msg_id=3)If the first messages.sendMessage query with msg_id=1 succeeds but the second with msg_id=2 fails, the query with msg_id=3 will fail with a MSG_WAIT_FAILED and will have to be resent as follows.
To recover the call queue, send the following new sequence of queries:
messages.sendMessage message=c (resending the old query with msg_id=3)Now assume the following, different sequence of queries:
messages.sendMessage message=amessages.sendMessage message=binvokeAfterMsgs msg_ids=[1, 2] (messages.sendMessage message=c)If either or both of the messages.sendMessage queries with msg_id=1 and/or msg_id=2 fail, the query with msg_id=3 will also fail with a MSG_WAIT_FAILED and will have to be resent as follows.
To recover the call queue, first of all wait for a response from the queries with msg_id=1 and msg_id=2.
Note how this is different from scenario 1, where we don't have to wait for a response from the previous queries: this is because in scenario 1, each invokeAfterMsg query was directly waiting for just one previous query, and any new query in the queue is chained to the previous invokeAfterMsg, thus any failure of any query with msg_id=N in the chain immediately blocked execution of all queries with msg_id <= N.
In this case, however, we're waiting for two messages at the same time, and the failure of the query with msg_id=1 does not prevent execution of the query with msg_id=2, and vice versa; thus when resending the query with msg_id=3 we have to either:
Wait for replies (errors or successes) for all of the queries mentioned in msg_ids, and then send the following new query:
messages.sendMessage message=c (resending the old query with msg_id=3)OR Wait for the receival of an RPC error for any of the queries mentioned in msg_ids, and then send the following new query:
invokeAfterMsgs msg_ids=([1, 2] - [$errId]) (messages.sendMessage message=c) (resending the old query with msg_id=3)I.e. resend the invokeAfterMsgs with all the initial msg_ids except for the one that failed.
It may be required to repeat the process if the newly sent invokeAfterMsgs also fails because the remaining query mentioned in msg_ids also failed (i.e. if queries 1 and 2 both fail at slightly different times).
Obviously the first option (waiting for replies (errors or successes) for all of the queries mentioned before proceeding) is the simplest.
An even simpler option is to always follow Scenario 1, never using invokeAfterMsgs and only using chained invokeAfterMsg calls, thus avoiding the use of this slightly more complicated recovery logic.
Important: if the helper methods invokeAfterMsg / invokeAfterMsgs are used together with invokeWithLayerN or other helper methods, invokeAfterMsg / invokeAfterMsgs must always be the outermost wrapper.
We recommend using gzip compression when calling methods to reduce the amount of network traffic.
The schema and constructor information are given in the protocol documentation.
Before transmitting a query, the string containing the entire body of the serialized high-level query (starting with the method number) must be compressed using gzip. If the resulting string is smaller than the original, it makes sense to transmit the gzip_packed constructor.
There is no point in doing the above when transmitting binary multimedia data (photos, videos) or small messages (up to 255 bytes).
By default, the server compresses the response to any request as well as updates, in accordance with the rules stated above. If the gzip_packed constructor is received as a response in rpc_result, then the string that follows must be extracted and uncompressed. Processing then continues on the resulting new string.