There is no easy answer to this question since it depends on many distinct factors. Not only must the work and computational resources required by the cryptanalyst be reasonable, but the amount and type of data required for the attack to be successful must also be taken into account.
One classification distinguishes among cryptanalytic attacks according to the data they require in the following way: chosen plaintext or chosen ciphertext, known plaintext, and ciphertext-only. (This classification is not particular to secret-key ciphers and can be applied to cryptanalytic attacks on any cryptographic function.)
A chosen plaintext or chosen ciphertext attack gives the cryptanalyst the greatest freedom in analyzing a cipher. The cryptanalyst chooses the plaintext to be encrypted and analyzes the plaintext together with the resultant ciphertext to derive the secret key. Such attacks will, in many circumstances, be difficult to mount but they should not be discounted. As Merkle and Hellman have remarked [MH81], a chosen text attack can in some ways be viewed as a "certificational weakness" in a cryptosystem.
A known plaintext attack is more useful to the cryptanalyst than a chosen plaintext attack (with the same amount of data) since the cryptanalyst now requires a certain numbers of plaintexts and their corresponding ciphertexts without specifying the values of the plaintexts. This type of information is presumably easier to collect.
The most practical attack, but perhaps the most difficult to actually discover, is a ciphertext-only attack. In such an attack, the cryptanalyst merely intercepts a number of encrypted messages and subsequent analysis somehow reveals the key used for encryption. Note that some knowledge of the statistical distribution of the plaintext is required for a ciphertext-only attack to succeed.
An added level of sophistication to the chosen text attacks is to make them adaptive. By this we mean that the cryptanalyst has the additional power to choose the text that is to be encrypted or decrypted after seeing the results of previous requests.
The computational effort and resources together with the amount and type of data required are all important features in assessing the practicality of some attack.